Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Status: EXPERIMENTAL

Stated here rather than only per-release, because a reader arriving at a specific version needs it as much as one reading the top.

This package has not run in production. While it is 0.x the API may change without a major version. Coverage is uneven by design: fakes and live public endpoints are well covered, order placement and authenticated flows are not.

Whenever an endpoint moves to :proven, the entry that does it states the evidence — what was run against the live venue, and when. "Marked proven" with no evidence is not an acceptable changelog line.

[Unreleased]

Fixed

• BREAKING: capabilities/0 declared has_staking: false (the default — the field was never set) while stake/3 and unstake/3 were already :experimental and genuinely reach Coinbase Prime. Found by dp_exchange_core's conformance-coverage audit, which added a cross-check (Capabilities.new/1, assertion 2 in the shared suite) requiring has_staking to agree with the six staking endpoints it summarises — the same rule already applied to supports_order_preview/supports_order_replace. Against Core 0.1.71 this package's own contract suite failed 23 of 42 tests, every one of them the same ArgumentError raised inside capabilities/0 itself, not 23 independent defects. Confirmed rather than assumed which side was wrong: DpExchange.Coinbase.Prime's nine endpoints are real paths against api.prime.coinbase.com, signed with Prime's own HMAC scheme, reached from stake/3, unstake/3 and this module's own query_transaction_validators/3, claim_rewards/4, staking_status/4, unstake_status/4 and preview_unstake_wallet/6 — not documentation describing a capability nobody wired up. has_staking is now true, at the same :experimental maturity as the rest of this :0.x package: the paths are read from Prime's own documentation, not probed live — this repository holds no Prime credential, and D7 tier 4 (money-moving) is answered in production by a consumer, never by a test here. measured_against says so explicitly. This is a behaviour change for a consumer routing on has_staking: it used to read false and now reads true, honestly, for a venue that has always been able to move money into and out of a staked position through this package. get_staking_rates/1, get_staking_balances/1, get_staking_rewards/1 and get_staking_history/1 stay :unsupported in venue_does_not_serve/0 — that reasoning is about their own shape (no published rate schedule, a status endpoint naming one wallet rather than every position, a rewards claim being a write not a report, no history endpoint at either scope) and has_staking becoming true does not reopen it. The same defect class already fixed in dp_exchange_gemini (has_staking: false/supports_margin: false beside six staking and three margin endpoints already :experimental).

• BREAKING: three defects in level2's unsubscribe-before-subscribe reconcile, all found by re-tracing the mechanism as a whole rather than as the sequence of fixes that built it, none caught by the existing suite. (1) A vanishing shard's stranded unsubscribe could be recorded and discarded in the same reshard/1 call — drop_unwanted_shards/3 dropped state.pending_unsubscribes[key] unconditionally for every shard no longer wanted, including one strand_unsubscribe/7 had just populated moments earlier, or one an in-flight deferred reconcile would populate a moment later — silently contradicting the notice text's own promise that it would be "picked back up on the next unconditional resubscribe cycle." (2) An ORDINARY reconcile (subscribe/2, unsubscribe/2, update_symbols/2) never consulted state.pending_unsubscribes at all — only the 60-second :resubscribe tick did — so a shard carrying one unconfirmed release could accept a plain additive subscribe for an unrelated new symbol on the very next ordinary call, the "quiet overflow" DpCryptoManagement's own probe 2 describes, reached through the path this whole design exists to close. (3) isolate_crashed_shard/5's reopen and an ordinary reconcile's own recovery of the same now-missing shard key could race: the deferred {:open_shard, _, _} handler always opened a fresh socket unconditionally, so whichever attempt's put_in ran last won state.shards[key]'s slot and the other's socket — still alive, still linked, still holding a live venue subscription — was leaked, never referenced by this module again. drop_unwanted_shards/4 now only drops a vanishing shard whose release is confirmed clear; reconcile_shard_in_place/8 folds pending_unsubscribes into what it treats as "current" on every reconcile, not only the 60-second cycle; handle_info({:open_shard, _, _}, _) now no-ops if the key already exists rather than opening a second socket. Breaking in the narrow sense that a shard whose vanishing release is still pending, or whose newly-added symbol was withheld behind one, now behaves differently (correctly) than before — no public API changed.

• A crash of Feed or Socket printed the CDP api_key/api_secret pair in cleartext, in OTP's own crash report. Both processes hold :credentials for their entire lifetime — Feed to keep resharding and resubscribing, Socket to sign every authenticated subscribe/4 — and both stored it as a bare map field in GenServer/ WebSockex state. OTP's default crash report prints a process's state in full on termination; a plain map prints every key it holds, secrets included. Verified by crashing an equivalent process holding %{api_key: "...", api_secret: "..."} as a bare state field and reading the resulting log line back — the key pair came back in cleartext. Process.flag(:sensitive, true) was tried as an alternative and does not help: the same crash, with the flag set, printed the same cleartext state; it disables tracing, not crash-report formatting. Now both processes wrap the pair in DpExchange.Coinbase.Credentials, a struct whose Inspect is derived with except: naming both fields, at the point credentials enter state — nowhere else in either module changes, because a struct is a map and Auth.jwt/2's %{api_key: k, api_secret: s} = credentials still binds the real values inside the one function that has to sign with them. Re-verified against a real crash of the new shape: the log line now reads credentials: #DpExchange.Coinbase.Credentials<...>. See Credentials's moduledoc for the full mechanism, including why it also closes a second leak (a FunctionClauseError's printed argument list goes through the same Inspect protocol as a crash report's state).

• A shard's socket crashing took the whole Feed down with it, silently discarding every subscription this feed had ever been given. Socket.start_link/1 runs inside Feed's own handle_call/handle_info (get_socket/1), which links every shard's socket to Feed the way start_link always does. Feed never called Process.flag(:trap_exit, true), so an abnormal socket exit — a decode bug raising inside a WebSockex callback, or anything else that kills the socket pid — sent an untrappable EXIT signal along that link and crashed Feed too. DpExchange. Coinbase.Supervisor then restarted Feed from the static opts it was given at tree-start, which never carry a consumer's later subscribe/2 calls: one shard's bug cost every symbol this feed was ever asked for, not just the shard that broke. Found by a 2026-09-07 supervision audit — proven by linking a real process into a running Feed the way get_socket/1 does and killing it with Process.exit(pid, :kill) (not :normal, which a non-trapping process ignores), which crashed Feed before this fix and does not after.

  Feed now traps exits and isolates a crashed socket to the one shard it belonged to: only that shard's symbols lose coverage, only the data_kind() that shard's channel carries is cleared from coverage/1/coverage_by_kind/1 (a level2 crash no longer erases a symbol's still-healthy ticker quote), a :link_down Core.Notice reports the crash, and the shard reopens immediately rather than waiting out the next :resubscribe tick (up to 60s by default). Every other shard, on other sockets, is untouched.

• BREAKING: supported_order_types and supported_time_in_force were both [] while {:place_order, 3} is :experimental and Rest.order_configuration/1's own @configurations cross-product builds three order types across four time-in-force values. Capabilities.new/1 validates the contents of these lists but never that a venue with an active place_order/3 declared anything, so two empty lists passed every check. Now supported_order_types: [:market, :limit, :stop_limit] and supported_time_in_force: [:ioc, :fok, :gtc, :gtd] — read directly off @configurations, not invented. :stop and :post_only are deliberately absent: the venue's cross-product has no :stop entry (only :stop_limit) and Advanced Trade's order endpoint has no post-only flag. Found by a cross-package audit; dp_exchange_robinhood defaulted the same two fields the same way for the same reason.

• child_spec/1 did not declare type: :supervisor, so OTP defaulted it to :worker — which also defaults :shutdown to 5_000ms instead of :infinity. A consumer terminating this child gave the whole nested tree (socket shards, rate limiter, and everything under them) only five seconds to shut down gracefully before :kill, rather than letting it unwind on its own terms. Invisible to any single-package review, and found only by diffing child_spec/1 across all five venue packages against each other; dp_exchange_schwab was the only one that already declared it.

• get_top_of_book/2 answered a missing local credential with {:refused, :missing_credentials} — in the real Rest client and in Fake alike — even though the credential never left this process and nothing at Coinbase ever saw a request to decline. DpExchange.Core.Venue's own moduledoc reserves :refused for the venue's own permanent word about a request it actually received; a locally-detected precondition is an :error. Found by a cross-package audit: Gemini, Robinhood and Schwab's real facades already used {:error, {:missing_credentials, venue}} for this exact condition — this package and Schwab's Fake (see that package's own changelog) were the two hold-outs. Now {:error, {:missing_credentials, :coinbase}}, matching the rest of the family.

• A credential that could not sign produced an unauthenticated request that was actually sent — on write endpoints, place_order/3 included. Auth.rest_headers/4 was written as Core.HttpClient's 4-arity auth hook, which may only return a header list and has no way to say "abort, do not send". So on a signing failure it returned the content-type header alone and documented that "the caller decides whether an unauthenticated request is acceptable". No caller ever decided: Rest.request/5 and Rest.json_request/5 both handed the result straight to HttpClient.request/5 without checking whether an Authorization header came back.

  For a public market-data GET that was harmless — Coinbase serves those anonymously. For json_request/5 there is no public path: every caller of it is a write, so a malformed api_secret turned a live order into an unauthenticated POST that the venue answered with an opaque 401, which reads as a credential problem at Coinbase rather than a malformed key here. That is precisely the failure the function's own moduledoc existed to prevent, produced by the mechanism documenting it.

  Auth.rest_headers/4 now returns {:ok, headers} | {:error, reason} and both request paths gate on it with with, the way every other venue package in this family already gated on its own Auth.headers. The nil-credentials branch is unchanged and is not this case: a call made deliberately without credentials still takes the public path.

  The test suite was green throughout, because the fixtures used an api_secret of "-----BEGIN EC PRIVATE KEY-----" — unparseable, and therefore never signing anything. Every fixture now carries a real 32-byte Ed25519 seed, so the suite exercises the signing path it was previously bypassing.

• Auth.jwt/2 raised KeyError on credentials it could not read, rather than refusing by name. It read credentials.api_key unconditionally, so nil, %{} or a map assembled with a typo'd key crashed inside signing — reachable from every write endpoint through Rest.json_request/5, which carries no nil guard of its own, and from Feed's unattended alias-map fetch. It now answers {:error, {:missing_credentials, :coinbase}}, the shape the other four venue packages in this family return for the same condition.

• 4xx statuses were recovered by searching the error message for "404", which was wrong in both directions. Core.HttpClient's own moduledoc names that exact expression as the reason its raw_status: true option exists:

  • False positive — any 4xx whose body contained "404" (an order id, a price, an embedded vendor code) became {:refused, :not_listed}: permanent, never retried, for something that may have been transient.
  • False negative — a 400, 401 or 403 contains no "404", so a bad request or a rejected credential fell through to {:error, message} and read as possibly-transient. A rotated key was retried instead of refused.

  Rest now passes raw_status: true and matches the status exactly, as the other four venue packages already did. 404 remains {:refused, :not_listed}; 400/401/403 are now {:refused, {:venue_error, status, message}} carrying the venue's own words.

  One existing test asserted {:error, _} for a rejected credential while its own name — and test_connection/2's moduledoc — both said refusal. Both were right; the assertion was pinning the bug.

Changed

• capabilities/0 now declares authenticated_streamable: [:order_book]. It was left at its [] default, which reads as "no streamed kind here needs a credential". That is not true: Socket's @authenticated_channels names level2, and Feed subscribes ["ticker"] without credentials and ["ticker", "level2"] with them — so an anonymous consumer gets quotes and no book. A host asking whether it needed a credential before it could stream book data was told no, and would have learned otherwise from a book stream that never arrived.

Added

• level2_pairs_per_socket is now a supervision option, not only the internal @default_level2_pairs_per_socket constant it defaults to:

  children = [{DpExchange.Coinbase, credentials: my_credentials(), level2_pairs_per_socket: 25}]

  Default is unchanged — 30, DpCryptoManagement's own live bisection against the real venue on 2026-09-06 (n = 6/12/25/30 accepted, n = 31/35/50/100 refused; adopted as the constant in 9139881) — the measurement is not being re-opened, only made adjustable. A value below 1, or a non-integer, fails Feed.start_link/1 at start with an ArgumentError rather than being coerced. A value above 30 is honoured, not capped — with a Logger.warning naming the measured ceiling, its date, and the concrete risk: an oversized level2 subscribe is refused by the venue WHOLESALE, losing that shard's entire coverage — and the socket SURVIVES the refusal (measured 2026-09-07, see the "corrected claim" entry below), which makes it quieter than a disconnect, not louder: no link-down, no reconnect, just a shard that silently never delivers. Capping it would have quietly defeated the option's own purpose, which is letting a consumer absorb a venue-side ceiling change without a package release; this repo has no way to verify such a change itself (tier-3, authenticated, live probing is out of scope for this repo — see the testing strategy). The default is deliberately not shrunk for headroom: 30 is the actual, located boundary, not a value merely bounded from below, and a consumer wanting margin below it can pass a smaller value themselves. ticker's own 100-per-socket size gets no equivalent option — it has no known ceiling to tune against. See feed.ex's own moduledoc, "level2_pairs_per_socket — a supervision option", and usage-rules.md for the full reasoning (DistortionPoint/dp-exchange-core issue #22).

• shard_spacing_ms is now a supervision option, not only the internal @default_shard_spacing_ms constant it defaults to — the delay between opening each successive shard's socket, whichever channel it carries, and the same delay reconcile_shard/7 now applies when reconciling several already-open shards in one update_symbols/2 call:

  children = [{DpExchange.Coinbase, credentials: my_credentials(), shard_spacing_ms: 1_000}]

  Default is unchanged — 5,000ms, inherited from the reference fix this package replaced, the same way @pairs_per_socket (100) is. A negative value, or a non-integer, fails Feed.start_link/1 at start with an ArgumentError, the same shape as level2_pairs_per_socket; 0 is NOT refused — Process.send_after/3 accepts it without complaint, so there is nothing mathematically broken about it, only extreme (a connect burst, the exact hazard this whole file otherwise staggers to avoid).

  Coinbase's own Advanced Trade rate-limits page states WebSocket connections are limited to 8 per second per IP, which converts directly into a floor of 125ms (ceil(1_000 / 8)) on this package's own connects. A value below that floor is honoured, not refused — this package cannot verify whether a faster pace is safe for a given consumer's own network position — but it logs a Logger.warning naming the documented floor, its source, and the concrete risk: connects tighter than the venue's own stated per-IP rate risk the connect-burst resets this package's own moduledoc opens with. See feed.ex's own moduledoc, "shard_spacing_ms — a supervision option", for the full reasoning, including why the 5,000ms default is left unmoved here even though it is roughly forty times more conservative than the documented floor requires (docs/design/ideas/shard-spacing-headroom.md records that as a non-blocking discovery, not acted on).

  This option exists first for this package's own test suite: five sharding tests in feed_test.exs existed to prove staggering happened and could previously only do that by waiting out the real production delay — 45 of the suite's roughly 51 seconds, across five tests. They now inject a small value and prove the same relative ordering and that a real delay was applied, rather than sleeping through the production default.

Changed

• BREAKING: Socket no longer maintains a level2 order book. An update frame now delivers dp_exchange_core's new Types.OrderBookDelta, not a full Types.OrderBook. A consumer that used to receive a rebuilt Types.OrderBook on every l2_data frame — including a single-row update — now receives a Types.OrderBook only on snapshot (once per subscribe/resubscribe) and a Types.OrderBookDelta on every update: the venue's own changed rows, in the venue's own order, both sides interleaved exactly as the frame carried them, in a flat levels :: [{side, price, quantity}] list rather than split into bid_levels/ask_levels. A consumer wanting a maintained book now builds and holds it itself. This is not presented as a performance improvement — it is the removal of state this package was never supposed to hold. See dp_exchange_core's docs/design/closed/2026-09-06_stop-maintaining-books-in-packages.md.

  Why: holding the book cost 65–110 ms per delta at the book size DpCryptoManagement measured live for BTC-USD (~22,800 bid / ~21,100 ask levels, issue #22), later optimised to ~6.6 ms — but that work ran on the same single-threaded process responsible for WebSockex.send_frame/2, so a socket busy rebuilding a book it was never asked to keep could not service its own sends, which is the :send_timeout behind issue #22. Maintaining state this package was not supposed to hold is what broke the connections it was supposed to keep; this change removes the work rather than making it faster a second time.

  A quantity of zero still means the level ceased to exist, not a price of zero — carried through completely unresolved now, since resolving it would itself be state-keeping.

  Reconnect reconciliation is now the consumer's job, not this package's. A dropped and resumed connection does not promise the deltas after it are contiguous with the deltas before it. subscribe_notices/1's existing :link_down/:link_up pair brackets where a gap may fall; neither that notice nor anything else reconstructs a missing delta. The correct response to :link_up is to re-pull get_order_book/2 (unaffected by this change) or accept the venue's own fresh snapshot on resubscribe — not to keep applying deltas across a gap. Coinbase's level2 channel publishes no book sequence number, so :sequence on both types is always nil here.

  Socket's books state, apply_book_event/3, apply_book_row/2, update_level/4, remove_level/3, deliver_book/3 and price_key/1 are all gone, along with bench/order_book_resort.exs, which benchmarked work that no longer exists. Feed.payload_kind/1 now maps %Types.OrderBookDelta{} to :order_book, the same data_kind() a full %Types.OrderBook{} gets — coverage_by_kind/1 answers "is book data arriving", not "in what shape", and the struct type itself already tells a caller which shape it is holding.

  Also dropped, deliberately, as part of the same change: the exact-scaled-integer precision check price_key/1 used to enforce (refusing a price with more than 8 decimal digits) existed only to support the :gb_trees ordering key that mechanism needed — it was never an independent business rule. Sorting a snapshot's own rows via Decimal.compare/2 needs no such key and has no rounding step to guard against, so a price at any precision the venue sends now passes through unchanged, the same as every other decimal field this module decodes. Likewise, two numerically-equal, differently-scaled prices in one snapshot ("1.5" and "1.50") are no longer folded into one last-write-wins level — that folding was an accidental side effect of the old map-keyed implementation's own key, never a documented venue behaviour (unlike the 8-decimal quote_increment finding, this had no live measurement behind it), and silently choosing a winner between two rows is itself the kind of substitution this family refuses. Both rows now pass through as the venue sent them.

Added

• Five of Coinbase Prime's nine staking endpoints are now reachable from the facade — dp_exchange_core's new "internal wiring" conformance assertion (assertion 16) caught them as built and never called from this package's own lib/. DpExchange.Coinbase.query_transaction_validators/3, staking_status/4, unstake_status/4, claim_rewards/4 and preview_unstake_wallet/6 now delegate straight to the matching DpExchange.Coinbase.Prime function, the same pattern stake/3 and unstake/3 already used for the other four. None of these map onto a dp_exchange_core.Venue callback — staking_status answers a narrower question than get_staking_balances/1 would, claim_rewards is a write where that callback wants a read, and query_transaction_validators/preview_unstake_wallet have no generic analogue at all — so they are Coinbase-specific facade functions, the same shape as the existing futures and portfolio extras (list_futures_positions/1, get_portfolio_breakdown/3, and friends). venue_does_not_serve/0's own documentation already claimed these were "reachable as Prime.X"; that claim is now true through the facade as well, not only by reaching past it into an internal module.

• coverage_by_kind/1 implemented — Coinbase is the motivating case for dp_exchange_core 0.1.48's new optional callback. coverage/1 answers "is anything arriving for this symbol" by counting any payload at all, so a level2 book update counted identically to a ticker quote. That blindness is not hypothetical: level2 delivered upward of 11,000 frames across 406 subscribed symbols while ticker stayed dark on all but a handful, and coverage/1 still answered :stream for all 406 — correct by its own definition, and exactly why DpCryptoManagement's issues #20 and #22 stayed unpinned for days.

  Feed's delivering map now tracks %{symbol => %{kind => timestamp}} instead of a bare timestamp, keyed off which Core.Types struct actually arrived (%Types.Quote{} → :quotes, %Types.OrderBook{} → :order_book) — never off this venue's own channel names, which stay internal. coverage_by_kind/1 on DpExchange.Coinbase and DpExchange.Coinbase.Fake both satisfy the union invariant dp_exchange_core's conformance suite now checks whenever a venue exports this callback: the symbol keys across every kind exactly match coverage/1's own keys, and every kind reported is one capabilities().streamable declares. The fake reports everything under :quotes only, honestly — subscribe/2 never synthesises an order book, and claiming :order_book coverage it cannot back would be the "differently capable" divergence this fake's own moduledoc forbids.

  Bumped dp_exchange_core from ~> 0.1.36 to ~> 0.1.48 to pick up the callback.

• Fake wired to Core.FakeInjection — DpCryptoManagement's issue #14. Every function with a real success path (not an unconditional Venue.not_supported()) now checks a queued or always-set outcome first: get_price/2, get_top_of_book/2, get_historical_prices/4, get_order_book/2, get_trades/2, quantization/1 and close_position/3 support per-symbol targeting; every other real function (bulk reads, account/portfolio/conversion calls, order placement, cancel/get/list, previews and edits) supports whole-call injection. subscribe/2, unsubscribe/2 and update_symbols/2 are deliberately not wired — each takes a symbol list in one call, which whole-call injection cannot express partial failure for; neither is coverage/1 or subscribe_notices/1, both local bookkeeping that always succeeds by construction.

  No credential-bypass mode here. Unlike DpExchange.Robinhood.Fake (the reference implementation), this fake has no central credential check to bypass — most functions never inspect credentials at all, an existing gap this wiring does not change. See docs/design/2026-09-04_webull-sharding-and-fake-injection.md §3.6/§3.7 in dp-exchange-core.

• get_market_overview/1 and list_instruments/1 are implemented — DpCryptoManagement's issue #10. Both sat behind Venue.not_supported(), one filed as a genuine venue absence (@venue_does_not_serve) with no per-item comment explaining why, against get_symbols/1 already calling the exact bulk endpoint (/products//market/products) that carries all of it. Live-verified: the response Coinbase actually returns names price, price_percentage_change_24h, volume_24h, high_24h, low_24h, status and product_type per product, and get_symbols/1 kept only product_id. Both new functions read the same response get_symbols/1 already fetches — via a shared fetch_products/1 — rather than a second request.

Removed

• Feed.pairs_per_socket/0 and SymbolFormat.mapping/0 deleted — both were public getters over a module attribute with no caller anywhere in this package's own lib/, the other shape assertion 16 exists to catch. Feed's own moduledoc already says sharding details "never reach the facade" — @pairs_per_socket is consulted directly by shards/1, the function that actually shards, and stays; the accessor was consulted only by a test asserting the same number shards/1's own tests already prove through behaviour. SymbolFormat.mapping/0's doc claimed it existed "so the conformance suite can drive CanonicalPair with it," but dp_exchange_core's conformance suite calls only to_canonical_symbol/1 and to_exchange_symbol/1 (the actual behaviour callbacks) — it never called mapping/0, and neither did anything else outside this package's own test suite. Breaking, for the two functions removed; neither was part of the Venue behaviour or documented as consumer-facing in usage-rules.md.

Fixed

• Corrected claim: a single oversized level2 subscribe does not close the socket. It is refused wholesale, and the socket survives — which is worse to detect, not better. Same issue #22. The level2_pairs_per_socket warning and docs/reference/coinbase/level2-session-limit.md used to leave this open: the 2026-08-26 incident recorded the refusal closing the socket, but the 2026-09-07 cumulative-overage probe (see the entry below) showed a refusal that did NOT close it — and nobody had re-run the single-oversized shape since 2026-08-26 to know whether that distinction mattered. DpCryptoManagement ran it, 2026-09-07: one socket, Socket.subscribe/4 for n = 30 (accepted, books=30), then n = 60 and n = 120 (both REFUSED — rate_limited, books=0, nothing delivering), in both ascending and largest-first order; Process.alive?/1 stayed true through a 40-second drain after every refusal, no :DOWN on a monitor, either ordering.

  The refusal is wholesale (the whole shard's coverage is lost, not truncated to the first 30) and the socket survives (no disconnect at all) — the identical shape the cumulative-overage probe already showed, so the two readings this package could not choose between turn out to be the same behaviour. The 2026-08-26 incident record is kept, dated, as unexplained rather than overturned: this measurement could not reproduce it, and neither this package nor DpCryptoManagement knows why. The risk is restated as more dangerous to notice, not less: a closed socket announces itself through a disconnect and a reconnect; a refusal on a socket that stays alive announces nothing beyond the :rate_limited Core.Notice Socket's error_kind/1 already emits — liveness looks perfect, and the oversized shard simply never delivers. coverage/1 / coverage_by_kind/1 are the only things that reveal it, unchanged by this correction: both already report only symbols that actually delivered a payload.

  The consumer's first attempt at this probe hit a harness defect worth recording for future probing with this package's own Socket: tearing a socket down with Process.exit(socket, :normal) is ignored by a process not trapping exits, so "closed" sockets stayed alive and leaked into the next attempt's count; fixing that to :kill then killed the probe itself, because Socket.start_link/1 links the socket to its caller — Process.unlink/1 first was needed. Recorded, dated and attributed, in docs/reference/coinbase/level2-session-limit.md.

  lib/dp_exchange/coinbase/feed.ex's moduledoc (the level2_pairs_per_socket section and its validate_level2_pairs_per_socket!/1 warning text), usage-rules.md, and docs/reference/coinbase/level2-session-limit.md are all updated; no code behaviour changed, only the documented and logged claim. No new Core.Notice was added: Socket's existing :rate_limited notice already names the venue's own refusal message verbatim, which already carries this specific cause — a second notice would duplicate one already firing.

• Socket full-re-sorted BOTH sides of the maintained level2 book on EVERY l2_data frame, including an update changing a single price level — dp_exchange_core's docs/design/2026-09-06_order-book-resort-cost.md. Measured at the book size DpCryptoManagement reported live for BTC-USD (issue #22, ~22,800 bid / ~21,100 ask levels): one update frame cost 62–110 ms across repeated runs in this repo (bench/order_book_resort.exs) — a ceiling of roughly 9–16 book updates/second, maximum, on a socket a shard shares across up to 100 symbols. Stated as a hypothesis in the design, not a conclusion here: this is very likely a major part of why ticker starves whenever level2 is delivering broadly in #22, since the single-threaded socket could never idle long enough to service the ticker subscribe inside FrameSender's 5-second window — but that causal claim is only confirmed by this fix changing behaviour on their node, not by anything measured here.

  Each side's %{Decimal => Decimal} map is now a :gb_trees tree keyed by the price scaled to an exact integer (10^8 — verified 2026-09-06 against Coinbase's own public GET /api/v3/brokerage/market/products: the smallest published quote_increment across all 931 products is 0.00000001, 8 decimal places, and no product's own price field carries more precision than that either), carrying the original Decimal as the tree's value. Delivery is now an ordered traversal — Enum.sort_by/3 no longer runs anywhere in the per-frame path. Measured against the same book size, repeated runs in this repo: the same one-update-frame cost fell to 0.9–2.9 ms (roughly 345–1075 updates/second, maximum), and the isolated bids-only sort/traversal fell from 49–83 ms to 0.5–6.2 ms. Run mix run bench/order_book_resort.exs to reproduce on any machine.

  A price that cannot be represented exactly at that scale is refused and reported through the same :data_quality notice path as any other unparseable row, rather than rounded — no real Coinbase price has needed this path so far, but the family's own rule against silently substituting a nearby value applies here too. One behaviour is deliberately NOT identical to the map-keyed implementation: two numerically-equal, differently-scaled price strings ("1.5" and "1.50") used to become two map entries — two "levels" at one price, because %Decimal{} structs compare unequal by field even when Decimal.equal?/2 says they are the same number — and now collapse into one, last-write-wins. That was a latent defect in the map-keyed version, found and fixed here rather than a behaviour changed as a side effect; everything else observable — order, Decimal values, count, the Core.Types.OrderBook struct itself — is unchanged, and is asserted so against a reference reimplementation of the old approach in socket_test.exs.

• The alias-map fetch added for issue #22 was throttled by the caller's OWN rate limiter at boot and never retried, disabling attribution for the life of the process — DpCryptoManagement's issue #26, a regression in a fix this package shipped. Feed's default alias_map_source called Rest.get_alias_map/1 without rate_limit_blocking: true, so it went through fail-fast check/3 instead of blocking acquire/3. The fetch is scheduled off the first subscribe/3, which for any real consumer is boot — the single most contended moment for their own limiter (universe discovery, catalogue reads and market overviews all landing at once) — so it was scheduled into exactly the window most likely to throttle it. One throttled call there, on code that never retried, was permanent: state.alias_map stayed %{} for the life of the process. Measured live: 406 pairs requested as -USDC, delivered as -USD, overlap 5 — coverage_by_kind/1 and the consumer's own tracker each reporting a truthful, and wildly different, count, precisely the situation the issue #22 fix existed to end.

  This is the third instance of one family-wide pattern — dp_exchange_robinhood's issue #16, this package's own issue #23 sweep, and now this: a background call with nothing waiting on it, failing instead of waiting, while Core.HttpClient's own error message names the fix in its text ("callers that can wait should set rate_limit_blocking: true"). The issue #23 sweep audited every REST call site in this package and missed this one, because the alias-map fetch's own HTTP path was mistaken for its sibling WebSocket resubscribe path, which genuinely has no rate-limited replay to default. That recurrence is worth more than any one of the three individual fixes: the same shape of gap keeps landing at the one call site nobody thought to re-check.

  Three changes, all in Feed:

  1. rate_limit_blocking: true is now set unconditionally on the fetch, by a new default_alias_map_source/2 that also forwards :limiter, :plug, :timeout, :retry_attempts, :retry_delay and :weight from start_link/1's own opts — the same allowlist shape Rest's own request pipeline uses.
  2. Blocking removes the self-throttle as a failure mode but not every failure: the limiter's own bounded wait can still time out. transient_alias_map_failure?/1 classifies that one case as worth retrying — the identical request can reasonably succeed once the bucket has drained further — and treats everything else (an unrecognised response shape, a refused request, an unclassified reason) as permanent, matching transient_subscribe_failure?/1's own default-to-permanent stance one section up. Retries are bounded (@max_alias_map_retries, backed off by @alias_map_retry_delay_ms, both overridable for a test's benefit) rather than looped forever.
  3. The degraded-attribution notice was unreceivable by construction, also part of issue #26. The fetch is scheduled from subscribe/3; a consumer calling subscribe_notices/1 afterward — the ordinary sequence — could register only after the fetch had already failed and fanned out to zero subscribers. A notice announcing a persistent degraded state that can only ever fire in the one window before anyone could be listening is worse than no signal: it looks like a working alarm that never rings. Fixed by replay, not by moving the emission earlier (this file already learned that moving-the-window lesson once, with next_resubscribe_delay/1's per-tick storm): alias_map_status and the reason that produced it now persist in state, and subscribe_notices/1 replays the identical notice to a newly-registered subscriber whenever it finds the state already :unavailable — once per new registration, never on a timer, and never re-sent to a subscriber who already has it.

  New tests in feed_test.exs drive the real fetch pipeline end-to-end for the first fix — a real, named Core.DefaultRateLimiter with its sole token already spent, behind a fake :plug response, proving rate_limit_blocking: true actually reaches Core.HttpClient rather than merely surviving being typed into an allowlist — plus the transient-retries-and-succeeds, permanent-fails-once, bounded-exhaustion and late-notice-subscriber cases, all against the default alias_map_source codepath or a controlled stand-in, none against a guessed Process.sleep/1 duration.

• An empty pricebooks array from /best_bid_ask was read as the venue naming a product not listed, and there is no evidence this venue has ever said that this way — audited alongside DpCryptoManagement's issue #25 (dp_exchange_robinhood's confirmed instance of the same substitution). get_top_of_book/2's {"pricebooks" => []} clause turned a 200 with an empty array into a permanent {:refused, :not_listed} — permanent because Core.PollingFeed reports a refusal once and never retries it.

  Probed live 2026-09-06 against the closely related, unauthenticated /market/product_book (same pricebook data, one product per call instead of a batch): a product this venue has never listed answers 404 {"error":"NOT_FOUND","error_details": "valid product_id is required"}; a product it delisted but still recognises (/market/products/{id} still answers 200) answers a different 404 {"error":"NOT_FOUND","error_details":"no pricebook found"}. Neither is a 200 with an empty array, and no online product checked (923 listed, spanning the lowest-volume pairs) ever returned one either. This venue's own convention for "no book" is a distinguishable non-2xx statement. /best_bid_ask takes a list of product_ids and answers one pricebook per product it can — an ordinary batch-API shape is to omit an entry it cannot answer rather than fail the whole request, which collapses "never listed" and "listed but delisted" (two states the sibling endpoint tells apart) into one indistinguishable silence, and says nothing about a real, momentarily bookless product either.

  The empty-array clause now returns {:error, :empty_result} — retryable, the same shape a 500 already produces. A genuine venue statement (a 404, this venue's own convention) still reaches {:refused, :not_listed} through the existing classify/1 path, which this change does not touch. New tests in rest_test.exs and order_book_test.exs cover both: an empty array is retried, and a genuine 404 is still refused.

• A timed-out channel subscribe was logged and thrown away — no retry until the next 60s tick reproduced the identical failure, DpCryptoManagement's issue #22. FrameSender's own moduledoc says the whole point of turning a send_frame exit into {:error, :send_timeout} is that a slow socket becomes "a failed batch, which a caller can report and retry" — the retry half of that design was never wired into Feed. A level2 subscribe triggers a full per-symbol book snapshot; the socket is single-threaded and cannot service the next send_frame while decoding it, so firing ticker's subscribe @channel_spacing_ms later still landed inside that window on a 100-symbol shard and blew the hardcoded 5s send window. Dropped, forever, since nothing re-attempted it before the next resubscribe cycle recreated the same busy socket.

  Measured live across a real ~400-symbol consumer, five boots over roughly 5.5 hours — the exact inversion this predicts:

  state	quotes (ticker)	order_book (level2)
  broken (4 boots)	~5 / 406	~406 / 406, 11,000+ frames
  healthy (1 boot)	400 / 406	6 / 406

  When level2 got through broadly, its opening snapshot burst starved ticker; when the venue refused most level2 subscriptions outright (its own per-session stream limit — see the "sharded" section above), ticker had the socket to itself and got everything. A lone :send_timeout on a ticker subscribe was also observed directly in an earlier run.

  {:error, :send_timeout} and {:error, {:send_exit, reason}} are now retried — transient, since the identical request can reasonably succeed once a busy or briefly gone socket catches up. {:error, {:credentials_required, channel}} is not: no amount of waiting supplies a credential that was never given, and it fails loudly on the first attempt instead of looping. The backoff reuses @channel_spacing_ms rather than a second, independently guessed number for the same busy-socket wait, bounded to two retries (three attempts total) — the whole chain resolves in at most 24s, well inside even the 60s default resubscribe cycle, so it can never stack fresh frames against the unconditional re-issue. Exhausting the retries, and the permanent-error path, both now emit a Core.Notice of kind :coverage_change in addition to the existing log — a channel that never subscribed is exactly the invisible half-dead feed this issue is about, and a Logger.warning alone gave a consumer no facade-level way to see it. A socket that dies between attempts is re-checked, not assumed alive, and simply stops the chain rather than sending into a corpse.

  Deliberately unchanged: @channels order (level2 before ticker) and @channel_spacing_ms itself. Subscribing the lighter channel first is a plausible additional fix, but it is unmeasured and changing two things at once would make the next measurement uninterpretable — raised separately with the consumer instead.

• :rate_limit_blocking was unreachable on every REST call this package makes — family-wide gap, DpCryptoManagement's issue #23. Core.HttpClient.check_rate_limits/1 reads this option to choose acquire/3 (wait for capacity) over fail-fast check/3, and its own error message on a self-inflicted throttle tells a caller to set it — but no caller could, on this venue: Rest.request/5, Rest.json_request/5 and Prime.request_opts/1 all stripped it from their forwarded-options allowlist before it ever reached Core.HttpClient. The same defect (dp_exchange_webull's issue #23, dp_exchange_robinhood's issue #16) audited across the rest of the family; this venue was one of four still carrying it.

  All three allowlists now forward :rate_limit_blocking, proven with a recording rate limiter that records which of acquire/3 / check/3 was actually called — not merely that the keyword survives the allowlist. Not defaulted anywhere in this package, unlike dp_exchange_webull's Feed and dp_exchange_robinhood's Feed: this venue's own periodic resubscribe (DpExchange.Coinbase.Feed's unconditional 60s re-issue) sends WebSocket frames, not HTTP, so there is no rate-limited background replay here to justify choosing a default on a caller's behalf. A caller that wants blocking opts in explicitly.

• A resubscribe interval shorter than one re-issue cycle wedged the feed — including the 60s DEFAULT, past twelve shards. A cycle is not instantaneous: shards are staggered @shard_spacing_ms apart and each shard's channels @channel_spacing_ms apart, so the last frame goes out about (shards - 1) * 5_000 + 8_000 ms after the tick. If the timer re-fired before that, cycles overlapped, frames queued behind each other, WebSockex.send_frame/2 blew its window, and the Feed stopped answering calls entirely — :sys.get_state/1 timing out. A wedged feed is strictly worse than a late resubscribe.

  Found by DpCryptoManagement while running the diagnostic added in the previous release (issue #22): they set resubscribe_interval_ms: 5_000, below the 8s channel spacing, and lost the run to it. They reported it against themselves rather than against the option, which is how it got looked at properly — because checking it showed the same failure was reachable with no option set at all. The 60s default is shorter than the cycle span from twelve shards (1,101 symbols at @pairs_per_socket) upward, so a large enough consumer would have walked into it on defaults alone. The knob exposed a limit the default already had.

  The next delay is now derived from the shard count that actually exists at each tick, never from the configured value alone, and an extension is logged rather than applied silently — a diagnostic knob whose value is quietly ignored is its own trap. Nothing changes for any interval that was already comfortable.

• get_top_of_book/2 could never work without credentials, and the facade said otherwise — family-wide defect sweep, Coinbase B1. Unlike every sibling market-data reader in Rest, this one call is hardcoded to /best_bid_ask with no /market/best_bid_ask branch. Re-verified live 2026-09-05: authenticated is 401, and the public path a caller would expect by analogy is 404 — there is no public form to fall back to, so inventing one would have been exactly the "nearby substitute" this family refuses. Fixed by checking for credentials up front and returning {:refused, :missing_credentials} before sending anything, rather than surfacing the venue's 401 as an opaque error. DpExchange.Coinbase's moduledoc and capabilities/0's credential_benefit comment both claimed "the same market data is served publicly" without qualification — true of every other endpoint, false of this one — and both now name the exception. usage-rules.md carried the identical claim and is corrected the same way, since it ships inside the Hex tarball and is what a consuming agent reads.

• apply_book_row/2 silently dropped a level2 row it could not parse — family-wide defect sweep, Coinbase B2. Every other decode-failure path in Socket (deliver_ticker/3, deliver_book/3) reports a :data_quality notice through report_quality/2; this one returned the maintained book unchanged with no signal, against the module's own stated discipline ("a payload that did not parse is reported, not swallowed and not fatal"). Concrete cost: new_quantity: "0" is how the venue signals level removal, so an unparseable quantity silently ignored could leave a stale price level in the maintained book indefinitely with nothing indicating why. apply_book_row/3 now threads state through and reports a :data_quality notice for an unparseable price_level/new_quantity and for a row missing those keys entirely — the connection is still never torn down over one bad row.

• Socket.start_link/1 inherited WebSockex's own connect/recv timeouts by accident — family-wide defect sweep, Coinbase B4. No :socket_connect_timeout or :socket_recv_timeout was set, so WebSockex supplied its own defaults — measured in the vendored dependency, deps/websockex/lib/websockex/conn.ex:10-11: 6_000 ms connect, 5_000 ms recv. That matters specifically because Feed's open_shard/5 synchronous branch calls Socket.start_link/1 from inside a handle_call/3, and Feed's own @call_timeout is @frame_window_ms * 3 = 15_000 ms — a named, shared process, so every other consumer's subscribe/2, unsubscribe/2, update_symbols/2 and coverage/1 call queues behind that one call. The inherited defaults alone (6_000 + 5_000 = 11_000 ms) would burn roughly three-quarters of that budget on the TCP connect and the handshake recv alone, against an unreachable or black-holing venue, before a single subscribe frame is sent. The margin was never chosen; it was whatever the dependency happened to default to.

  Fixed by setting both explicitly at 3_000 ms each (6_000 ms total), chosen deliberately against Feed's 15_000 ms budget — leaving roughly 9_000 ms of the same call for the socket to send at least one subscribe frame (capped at Feed's own 5_000 ms @frame_window_ms) plus ordinary GenServer overhead. No failure semantics changed: start_link/1 still returns {:error, reason} synchronously exactly as before, so the synchronous-primary-shard design is unchanged bit for bit — only the margin after a slow or absent venue does. A caller passing either key in opts still overrides it. The merge is factored into a small @doc false connection_opts/1 so a regression test can pin both the defaults and the override precedence without opening a real socket.

• The venue rewrites an aliased product id on delivery, and streaming passed the rewritten id straight through — DpCryptoManagement's issue #22. Measured live 2026-09-05 against wss://advanced-trade-ws.coinbase.com: subscribing ticker to ["XLM-USDC", "AVAX-USDC"] — sent exactly as asked, both real, listed products — delivers every frame tagged XLM-USD and AVAX-USD instead; the venue's own subscription acknowledgement even echoes the rewritten names back ("ticker" => ["XLM-USD", "AVAX-USD"]), not the ones actually sent. This is the venue's own declared behaviour, not a guess: the same public, unauthenticated /market/products catalogue this package already reads for get_symbols/1 and list_instruments/1 names it directly — on this date, 112 of the first 114 USDC products carried a non-empty alias naming their -USD counterpart. On a settled DpCryptoManagement node running 0.1.17 with 406 pairs requested, this was the same defect wearing two faces: 174 of the 406 requested pairs delivered nothing under the name asked for, while 401 pairs never requested were decoded and stored under a name nobody subscribed.

  Fixed in Feed, not Socket: Socket still decodes and delivers under whatever product_id the venue actually sent, unchanged. Feed.handle_info({:dp_exchange, :coinbase, payload}, state) now resolves a delivered id against Rest.get_alias_map/1 — the venue's own declared relationship, fetched once, asynchronously, the first time subscribe/3 or update_symbols/2 runs (never per frame, never per subscribe; see Feed's moduledoc for why it is not read from init/1 or inline in the triggering call) — and delivers under every name in wanted that names the same market: the caller's own requested name, and its alias where the caller subscribed to that instead. A caller subscribed to both receives both, from one delivered frame. coverage/1 needed no code change to become honest, since it already reports whatever key delivery is recorded under.

  A catalogue that cannot be fetched degrades rather than guesses. A failed fetch delivers under the venue's own id — today's pre-fix behaviour — and reports exactly once, as a :data_quality notice naming the failure, that attribution is degraded and why. Munging -USDC into -USD was considered and rejected: it would be exactly the "nearby substitute" this family forbids, and wrong for any pair the venue does not alias — nothing here assumes the suffix relationship holds in general, and the fix reads only the venue's own alias field.

  Regression tests in feed_test.exs drive the proven mechanism directly — a subscribe to the alias form receiving frames tagged with the canonical form delivers under the alias form; coverage/1 lists what was requested; both names subscribed both receive one delivered frame; a catalogue fetch failure delivers under the venue's id plus the degraded notice, never a guessed mapping; the fetch happens once regardless of how many subscribes or delivered frames follow — and rest_test.exs covers get_alias_map/1 itself against a catalogue shaped like the live response captured while proving this.

• Supervisor's own rate limiter was configured from public_ceiling unconditionally, contradicting capabilities/0's own documented promise that credentials buy the higher ceiling. capabilities/0 states outright: "Pass credentials and this package uses the authenticated path, which has the higher ceiling." Rest's request paths honour that — but Supervisor's init/1 started DefaultRateLimiter from caps.public_ceiling (3 req/s) for every instance, credentialed or not, so a consumer supplying credentials got the documented 10 req/s from the venue and a third of that from this package's own throttle regardless — a mechanism silently disagreeing with the declaration it exists to encode, per this module's own moduledoc rule. limits/1 now reads opts[:credentials] (the same opts Feed reads it from) and configures the bucket from authenticated_ceiling when a non-empty credential map was given, public_ceiling otherwise. New tests in supervisor_test.exs prove both ceilings actually reach the running limiter, and that an empty %{} does not buy the higher one.

• A shard whose socket failed to open at all was silent at the facade — no Core.Notice, only a Logger.warning that never crosses it. Every shard beyond the first opens asynchronously; when its Socket.start_link/1 failed (a transient connect refusal, a timeout), handle_info({:open_shard, _, _}, _) logged and stopped, leaving those symbols silently absent from coverage/1 with nothing telling a consumer why — the exact "silent half-dead feed" this module's own moduledoc is about, and worse than the channel-subscribe case one section up, which already got a :coverage_change notice for the identical shape of fact (subscribed intent that did not become delivery). Now emits one, through the same notify_shard_open_failed/3 path.

  Worse: a shard that failed to open had no automatic recovery path at all. The unconditional :resubscribe tick only ever walked state.shards — a shard whose socket never opened is not a key in it, so the tick had nothing to re-issue for it, and the only thing that would ever reconsider it was a fresh subscribe/3 or update_symbols/2 call, which may never come for a consumer whose scope is stable after boot. retry_missing_shards/1 closes this: on every :resubscribe tick, any shard state.wanted still implies but state.shards has no entry for is retried on the same unconditional cadence an already-open shard's subscriptions are re-issued on, staggered past them by the usual @shard_spacing_ms. New tests in feed_test.exs cover both: the notice on a failed open, and the tick recovering a shard that never opened.

• Reconciling more than one ALREADY-OPEN shard in a single update_symbols/2 call dropped the stagger between them — only a brand-new shard's connect was staggered. reshard/1 computes position * @shard_spacing_ms for every shard beyond the synchronous primary and hands it to touch_shard/4, but reconcile_shard/6 (an existing, already-open shard) never received it — every already-open shard a single call touched had its level2 subscribe scheduled at the identical instant regardless of position. This is not the connect burst @shard_spacing_ms was written against (no new socket opens here), but a related hazard: Socket.subscribe/4 blocks THIS Feed process — via FrameSender, up to WebSockex.send_frame/2's 5s window — for as long as its target socket takes to acknowledge, and several such messages landing in this process's single mailbox together serialise into back-to-back blocking sends, capable of stalling coverage/1 and every other call to this Feed for as long as the slowest one takes. reconcile_shard/7 now receives and applies the same delay open_shard/5 already did. A new test in feed_test.exs proves it with two live sockets recording arrival time: touching two already-open shards in one call now delivers their first frames @shard_spacing_ms apart, not together.

• Fake's get_top_of_book/2 ignored opts[:credentials] entirely, making the fake MORE capable than the real venue on the one call where that gap matters. Rest.get_top_of_book/2 refuses {:refused, :missing_credentials} before sending anything when given none — /best_bid_ask has no public form on this venue, confirmed live (401 authenticated, 404 at the /market/... path a caller would expect). The fake answered {:ok, %Types.TopOfBook{}} regardless, so a consumer's test written without credentials would pass against the fake and refuse identically in production — precisely the silent "differently capable" divergence this module's own moduledoc says it exists to prevent ("Six were loud... Three were silent, and those are the ones this is designed against"). The fake now refuses the same way, gated on the same field. New tests in fake_test.exs and an updated one in fake_injection_test.exs (which previously called it with no credentials at all and got away with it) cover both branches.

• level2 and ticker shared one shard size, and the venue's own level2 ceiling is well under it — DpCryptoManagement's issue #22 continuing, not reopened. The ticker starvation fixed earlier in this file (the timed-out-subscribe retry entry above) is a separate, already-closed incident in the same file; this is a second, independent defect the sharding line above deliberately left unchanged pending a separate measurement ("Deliberately unchanged: @channels order... raised separately with the consumer instead"). That measurement is this entry.

  A consumer's real 406-symbol universe, on sixteen otherwise-healthy boots (0 send_timeout, 406/406 quotes coverage): shards/1 at the old shared @pairs_per_socket (100) split it [100, 100, 100, 100, 6], and level2's subscribe was refused on every 100-symbol shard — "too many L2 streams requested in a single session", 5,099 times — while the six-symbol shard's was not. coverage_by_kind/1 answered order_book: 6 throughout: exactly the tail shard's own count, pinning the cause on shard size rather than the connection, the alias fix, or ticker (which has no such ceiling and was unaffected on the same boots).

  No Coinbase documentation states a per-session level2 product ceiling — re-checked 2026-09-06 against the Advanced Trade channels reference, connection overview and rate-limits page (which states an 8-per-second-per-IP connect/message rate, not a subscription count), and the older Exchange product's separate rate-limits page (which states a different, inapplicable number: 10 duplicate subscriptions to the same product-channel pair, not the count of distinct products, and for a product this package does not speak). This package cannot narrow it by probing the venue either: level2 is authenticated, and this repo's own testing strategy draws tier 3 (authenticated, live) as needing credentials this repo must never hold — the same line that already keeps this repo off order placement.

  level2 now gets its own shard grouping, at its own, independent size — @level2_pairs_per_socket, 6 — rather than sharing @pairs_per_socket (100, still ticker's own size, unchanged) with ticker. 6 is not a rediscovered venue limit; it is the largest level2 subscription size this package has direct evidence the venue accepts, taken from the production numbers above (100 refused four times out of four, 6 accepted once out of one) rather than guessed at some unverified point between them. Every shard, either channel, now opens its own dedicated, single-channel socket — for the 406-symbol universe above, 5 ticker sockets (unchanged) plus 68 level2 sockets (ceil(406 / 6)), 73 total against 5 before, affordable now that removing in-package level2 book maintenance (the change above this one) cut per-frame decode cost roughly tenfold.

  Every new socket — either channel — is staggered on one @shard_spacing_ms sequence with every touched ticker shard ordered ahead of every touched level2 shard, so ticker's own boot-time coverage stays exactly as fast as before this change while level2's far more numerous shards ramp in behind it — for the 406-symbol universe, roughly six minutes for the last level2 shard, against a ceiling that previously never moved at all. @channel_spacing_ms — the wait between level2 and ticker sharing one socket — is deleted along with the shared-socket design it existed for; no socket carries two channels any more, so the busy-decoding hazard it guarded against cannot occur. @subscribe_retry_delay_ms keeps its value (still 8_000ms) on its own reasoning rather than borrowing from a constant that no longer exists.

  An adaptive shard size, driven down at runtime by the venue's own refusal so a wrong constant could never be silently wrong forever, was considered and not built: correctly telling a live shard's bookkeeping apart from a stale one the venue already emptied on refusal is real complexity with its own correctness risk (silently under- or double-subscribing a shard), and did not clear its bar against a fixed, evidence-grounded constant plus the safety net that already existed and needed no change — Socket's error_kind/1 already classifies "too many" as :rate_limited and reports it as a Core.Notice on every occurrence, and coverage_by_kind/1 already never marks a symbol covered for :order_book on subscribed intent alone. Both are verified unchanged by this fix. If 6 is ever also refused, a consumer with subscribe_notices/1 wired up hears about it exactly as loudly as any other refusal in this file, and lowering the constant is a one-line change rather than a runtime decision made silently.

  New tests in feed_test.exs pin the structural fix: a credentialed feed given a symbol count that fits in one ticker shard splits it into two level2 shards, with the ticker shard chosen as the call's synchronous primary; a credential-less feed never opens a level2-keyed shard at all. The 12-shard resubscribe-floor test is now a 13-shard one, and its expected numbers drop the deleted @channel_spacing_ms term — both mechanical consequences of this change, not new behaviour of their own.

• @level2_pairs_per_socket was 6 — a conservative lower bound, correctly labelled as one — and the real boundary is now measured: 30. Same issue #22, continuing. 6 was never a rediscovered venue limit; it was the largest size this package had any positive evidence for at the time, with 100 (refused) as the nearest known failure and nothing in between actually tried. This package still cannot bisect a live, authenticated level2 session itself — that is tier 3, a line this repo does not cross for any endpoint. DpCryptoManagement can, and did, on 2026-09-06 (issue #22): a fresh socket per attempt, never reused, against their real 406-symbol scope — n = 6/12/25/30 accepted, n = 31/35/50/100 refused, confirmed by interleaving two runs back to back and by a contamination check (re-running n=6 and n=30 immediately after a refusal run, both still accepted, ruling out probe-induced saturation as the cause). 30 is the largest value with positive evidence of acceptance, 31 the smallest with positive evidence of refusal — the boundary itself, not merely a known-good far below a known-bad. See docs/reference/coinbase/level2-session-limit.md for the full method and attribution.

  For DpCryptoManagement's 406-symbol universe this is 14 level2 sockets instead of 68 (19 total instead of 73) and roughly 70 seconds to full order_book coverage instead of about six minutes — the same staggered, ticker-first connect sequence as before, just markedly shorter because far fewer sockets need it.

  A second question this raises got checked directly, not left open by omission. The consumer's harness used a fresh socket per attempt specifically to keep cumulative session state out of its own result, which means it cannot say whether Coinbase's ceiling counts concurrently-held products or every distinct product a session has ever carried. Checked against this package's own code: reconcile_shard/7 used to subscribe a shard's newly-added symbols onto whatever socket that shard already had open, and a MapSet's enumeration order being a function of its current keys (not insertion order) means ordinary universe churn — not only a caller literally adding a symbol — routinely hands an already-open level2 shard products it has never carried before. At 6 this had enormous headroom; at 30, aimed at a now-exact boundary, it had none: the very first churn past a full shard could have pushed one socket's lifetime subscription count to 31 even though its concurrent membership never left 30. Fixed by replacing the socket instead of mutating it whenever a level2 shard's membership would grow (reconcile_shard/7's "level2" clause, replace_level2_shard/7, terminate_socket/1) — a shard that only loses symbols keeps its existing socket, since removal cannot grow that count, and ticker is unaffected, since it has no known ceiling to protect. This makes "should never be over the ceiling" hold unconditionally, concurrently and cumulatively, per socket, regardless of which of Coinbase's two possible countings turns out to be real.

  One axis is still genuinely open and was not fixed here. The unconditional 60-second resubscribe re-issues a shard's unchanged symbols to its already-subscribed socket forever, which would feed an attempt-counted ceiling if Coinbase has one — this is the same open question Feed's @default_resubscribe_interval_ms comment already named before this investigation, and this investigation did not close it. Two specific probes that would are recorded in docs/reference/coinbase/level2-session-limit.md, for whichever party next holds the credential to run them.

  DpCryptoManagement also reported a refusal that was not always a clean gate — some oversized subscribes delivered 1,300+ books alongside their rate_limited notice rather than refusing outright, once even at n=31, the smallest over-the-boundary value. This does not move the boundary (every n ≥ 31 refused, every n ≤ 30 did not) and has no explanation from either party; it is recorded, dated and attributed, as an unexplained venue characteristic in docs/reference/coinbase/level2-session-limit.md, not rationalised into a theory neither party has evidence for. This package should never itself trigger it — every level2 subscribe it sends carries at most 30 symbols by construction, before and after this fix — and coverage/1 / coverage_by_kind/1 need no change to stay honest if it ever does: both are built entirely from symbols that actually delivered a payload, and a Core.Notice never touches that bookkeeping.

  New tests in feed_test.exs pin both the number and the fix: 35 symbols (rather than 10) now produces two level2 shards at 30/socket; a level2 shard engineered to be missing symbols its own fresh chunking would include is replaced — old socket killed, new one recorded, target set intact — while one only losing symbols keeps its existing socket; ticker given the identical setup keeps mutating in place; the deferred (non-primary-shard) replace path is driven directly, including the stale-message guard and a replacement socket that fails to open.

• The socket-replacement fix directly above is superseded: the question it hedged against is now measured, and the hedge was more expensive than the answer required. Same issue #22, continuing. 9139881 replaced a growing level2 shard's socket outright because this package could not tell whether Coinbase's per-session ceiling counted concurrently-held products or every distinct product a session had ever carried — a real open question at the time, and a full reconnect (fresh snapshot, coverage gap) was the cost of not guessing wrong either way.

  DpCryptoManagement answered it directly, 2026-09-07: three probes, each on ONE socket, using raw Socket.subscribe/4 — deliberately not Feed/update_symbols/2, so the result is evidence about the venue's own accounting, not this package's dedup. The same 30 products re-sent ~60s apart 14 times (~13 minutes): all accepted, books=0 on every repeat — repeats do not accumulate, which also directly answers the "attempt-counting" open question from the previous entry (the unconditional 60-second resubscribe does not feed a cumulative counter, at any shard size). A different 30 products on the same socket without unsubscribing the first batch: cumulative=60 — REFUSED, but the socket stayed alive and the first batch kept delivering (27/30 still ticking); only the second, unreleased batch was rejected. Four batches of 30, each unsubscribing the previous batch first: all four accepted, fresh snapshot each time — 120 distinct products through one socket, never more than 30 live at once. The ceiling is 30 CONCURRENT products per session, not 30 over a session's lifetime, and unsubscribe releases budget the venue actually honours.

  reconcile_shard/7's "level2" clause and replace_level2_shard/7 are gone. reconcile_shard_in_place/7 now handles both channels, all three shapes a shard's membership can change (only losses, only gains, both), on the shard's EXISTING socket: removed is unsubscribed before added is subscribed, and the subscribe never goes out until the unsubscribe has returned :ok (retried on a transient send failure with the same bounded backoff a channel subscribe already gets — attempt_channel_reconcile/6, handle_unsubscribe_failure/8 — with added withheld, not sent anyway, if the retries exhaust). The order is load-bearing: probe 3 above works BECAUSE the departing batch's slots were freed before the arriving batch was requested; probe 2 is the identical operation in the other order and was refused, quietly — the socket stayed alive and already-flowing symbols were unaffected, with only the newly-requested ones silently missing. Shrinking before growing keeps a socket's transient concurrent count bounded by max(length(current), length(wanted)), never more than level2_pairs_per_socket, including mid-reconcile, not only at rest.

  What guarantee this package actually has is stated plainly, not assumed generous. Socket.subscribe/4 and Socket.unsubscribe/3 both block on FrameSender's synchronous WebSockex.send_frame/2 call, so ordering the two calls in code orders the two frames on the wire; TCP delivers one connection's bytes in the order they were written, and nothing in any probe has ever shown Coinbase reordering two frames on one connection. But Socket.unsubscribe/3's :ok means the frame was handed to the connection, never that the venue has finished releasing the departing slots — this protocol gives no acknowledgement frame for an unsubscribe to wait on. This package relies on frame order, not a confirmed venue-side state transition, and says so in feed.ex's own moduledoc ("unsubscribe before subscribe") rather than rounding the guarantee up. One gap stays open and disclosed: a permanently stranded unsubscribe (every retry exhausted) is not itself retried by the next resubscribe cycle, which only re-issues a plain subscribe — closing that would need a cross-cycle retry ledger judged not to clear its own complexity bar against how narrow the gap is.

  docs/reference/coinbase/level2-session-limit.md records all three probes, dated and attributed, and marks the cumulative-vs-concurrent and attempt-counting questions resolved. One thing is explicitly NOT resolved and is recorded as open, not decided either way: the level2_pairs_per_socket warning and the reference doc both used to state that an oversized subscribe "closes the socket and loses that shard's entire coverage" — probe 2 above shows a cumulative overage refusal that did NOT close the socket. The original 2026-08-26 incident recorded socket closure for what may be a different case (a single oversized subscribe, not cumulative overage), and the consumer has offered to test the single-oversized case specifically; until that runs, both observations are stated and the conflict is left open, and the warning text states the worse of the two outcomes as the risk to plan for rather than asserting it as certain. (That test ran the same day — see the "Corrected claim" entry above this section for the result: refused wholesale, socket survives, same as probe 2.)

  New tests in feed_test.exs replace the socket-replacement suite: a growing level2 shard reconciles on its existing socket (never replaced, never killed); a shard that both loses and gains sends the unsubscribe frame before the subscribe frame, on both the synchronous (primary-shard) and deferred (async) reconcile paths; a shard reconciling losses and gains together never reports more live products than its own shard size at any point a fake venue-tracking socket observes, including mid-reconcile; a transient unsubscribe failure is retried and the subscribe stays withheld until it succeeds; an unsubscribe that exhausts its retries withholds the subscribe entirely and reports a :coverage_change notice naming both the departing and withheld-arriving counts.

• The fix directly above left its own gap: a permanently stranded unsubscribe was dropped from this package's own bookkeeping, not merely from the venue. Traced by the coordinator against the actual code, not taken on the summary above's word: reconcile_shard_in_place/7 recorded state.shards[key].symbols = wanted unconditionally on both its sync and deferred clauses, regardless of whether the underlying unsubscribe ever succeeded. The WITHHELD added side already recovered — the 60-second cycle re-issues state.shards[key].symbols, which is wanted — but the STRANDED removed side did not: nothing re-issued the unsubscribe, and this module's own bookkeeping had already stopped tracking it as owed. Left alone this degrades into the exact failure this whole file exists to prevent — the venue's live count sitting at old ∪ wanted, eventually exceeding the shard's cap, refusing every later subscribe for that shard quietly (DpCryptoManagement's own probe 2 — socket alive, existing symbols still flowing, new ones silently absent) — and the added recovery makes it WORSE, not better, by continuing to push subscribes into a budget the stranded slots guarantee cannot fit.

  Closed using the machinery already there, per the coordinator's own instruction, not a new subsystem: state.pending_unsubscribes (%{{channel, index} => [symbol, ...]}) is written by strand_unsubscribe/7 whenever a reconcile gives up on releasing some removed symbols — the synchronous clause's single failed attempt (no retry chain of its own, so it strands immediately) and the deferred clause's exhausted retry chain alike — and read back by handle_info(:resubscribe, _) on the identical unconditional cadence retry_missing_shards/1 already uses to recover a shard whose socket never opened at all; that function's own comment states the governing principle this reuses word for word. Each tick now sends {:channel_reconcile, key, socket, channel, pending, symbols, credentials} per shard — the shard's own pending entry as removed, its WHOLE current membership (not a delta) as added — through the SAME attempt_channel_reconcile/6 an ordinary reconcile already runs, so the retried release precedes that shard's resubscribe on the same tick for the identical reason it precedes one anywhere else in this file. For the ordinary case (nothing pending) this costs nothing beyond what a plain Socket.subscribe/4 already cost, since unsubscribe_step/3 short-circuits an empty list without sending anything.

  clear_pending_unsubscribe/3 removes only the symbols a successful send actually covered (pending -- removed, not the whole key), so an unrelated, still-outstanding stranding on the same shard survives a different one clearing. strand_unsubscribe/7 is symmetric — it merges, never overwrites. drop_unwanted_shards/3 now also drops a shard's pending entry the moment the shard itself is dropped, since handle_info(:resubscribe, _) walks Map.keys(state.shards) specifically and would otherwise retry nothing for a key that no longer exists there — this is what keeps state.pending_unsubscribes bounded by currently-open shards carrying an unresolved stranding, not by this module's all-time history of failures. A symbol churned out (stranding it) and back into the same shard before the next tick is not a hazard either: added is always the shard's whole current membership, so the same reconcile's own subscribe puts it right back if still wanted, in order, costing at most one redundant frame pair.

  What was deliberately not built, and why, argued rather than assumed: correlating a stranding to a specific socket's own reconnect. A reconnect gets a fresh venue session, releasing every stranded slot on the old one whether this package notices or not — but Socket holds no shard-correlating identity in its :link_up/:link_down notices by design (its own moduledoc: "Socket holds no book to key by anything"), and WebSockex's reconnect keeps the same pid, so pid identity cannot substitute. Building that correlation would widen Socket's own contract for a case whose cost, left alone, is bounded: a stale entry retried against a fresh session sends one wasted unsubscribe (assumed idempotent-safe against an unsubscribed symbol, consistent with but not independently measured against this family's stated subscribe-idempotency assumption — labelled a gap, not hidden as a certainty) and is immediately followed, same reconcile, by a subscribe of the shard's whole current membership — self-correcting the same tick, never leaving the shard short a wanted symbol.

  transient_frame_failure?/1's real reach, stated rather than assumed generous: Socket.unsubscribe/3 builds no JWT and checks no credentials, so it has no analogue of a subscribe's {:credentials_required, channel} — its only two possible failure shapes, :send_timeout and {:send_exit, reason}, are both already classified transient. The non-transient cond branch in handle_unsubscribe_failure/8 therefore does not fire against the real venue today; "permanently stranded" is reached, in practice, only by EXHAUSTING the bounded retry chain against a socket that keeps failing to send without ever actually dying (Process.alive?/1 staying true throughout — a genuinely dead socket short-circuits every reconcile path before reaching this branch, matching every other dead-socket guard in this module). The non-transient branch is kept anyway, wired to identical behaviour, for the same "do not assume a shape that happens to hold today" reason handle_subscribe_failure/6 keeps its own — a two-branch cond reusing one existing helper, not new machinery for a case argued, not merely assumed, to be effectively unreachable.

  New tests in feed_test.exs: the resubscribe tick retries a shard's pending unsubscribe before it resubscribes and a successful retry clears it; a retry that also fails leaves the entry in place for the next cycle; a synchronous (primary-shard) unsubscribe failure is stranded rather than dropped; dropping a shard entirely also drops its own pending entry.

Documentation

• CLAUDE.md claimed this package parses Coinbase's cb-after / cb-before rate-limit headers; it deliberately does not, and never has. Those are pagination cursors, not rate-limit data, and Coinbase publishes no x-ratelimit-* or retry-after either — measured live 2026-08-28. A prior adapter's parser keyed off the cursor headers and returned three hardcoded constants labelled as measurements (remaining: 100, limit: 100, a reset time one minute out); porting it would have been exactly the fabrication this family refuses — recorded in docs/reference/coinbase/reconciliation.md §5.5, which CLAUDE.md's own claim contradicted. Corrected to state what actually happens: nothing is parsed, and Core.HttpClient's generic parser correctly answers nil.

• docs/reference/coinbase/endpoint-inventory.md still listed /best_bid_ask and /product_book as not implemented — family-wide defect sweep, Coinbase B3. Both were implemented and declared :experimental in capabilities/0 well before this release; the note was never updated when they shipped, which is part of why B1's missing public/private branch on /best_bid_ask went unnoticed. Both endpoints are now marked ✓ in the endpoint list, the stale "absent" note is corrected, and the newly measured fact from B1 is recorded where this file's other live measurements live: /best_bid_ask has no public form, verified live 2026-09-05, unlike /product_book, whose market/product_book twin is real and public.

• usage-rules.md's "Streaming" section never said which symbol a delivered frame carries, and never mentioned resubscribe_interval_ms at all — family-wide defect sweep, Coinbase B5. Both are consumer-facing behaviour a subscribing agent needs to act on correctly, and this file — the one that ships inside the Hex tarball and is not the README — was silent on both. The alias-attribution fix above changes what symbol arrives on every streamed frame; resubscribe_interval_ms has been a real Feed option, forwarded straight through from {DpExchange.Coinbase, resubscribe_interval_ms: ms}, since the resubscribe-wedge fix above added it, and neither fact was checkable from the shipped docs. Added two sections: one stating a delivered frame is tagged with the symbol the caller subscribed to, never the venue's rewritten alias, including the degraded-attribution fallback and its :data_quality notice; one documenting resubscribe_interval_ms's 60,000 ms default, how to set it, and that a value below one full re-issue cycle for the current shard count is silently clamped to the computed floor and logged rather than honoured.

• README's endpoint counts were stale. It read "46 are declared :experimental and 41 :unsupported" with "38" of those the venue's own absence. Run against the real capabilities/0 (mix run -e, 2026-09-05): 48 :experimental, 39 :unsupported, of which 37 are venue_does_not_serve/0 (the other 2 are @not_ported, get_funding/2 and get_contract_stats/2). Corrected to the measured numbers rather than re-guessed.

• frame_sender.ex's moduledoc claimed WebSockex.send_frame/2 has "no way to override" its 5-second timeout. The vendored websockex 0.5.1 exposes send_frame/3 with a timeout argument, so the claim was wrong. FrameSender.send/3 still calls the 2-arg form, so nothing about the actual timeout behaviour changes here — see the design doc's deferred section for why a longer timeout is a decision for later, not a drive-by alongside this correction.

• Every ticker frame from the real venue failed to decode — 0 Quotes delivered, ever, against live Coinbase, for the entire life of this package. Surfaced while chasing DpCryptoManagement's issue #22: a live test against 60 non-aliased, canonical -USD symbols captured 500+ consecutive data_quality notices and zero Quotes in a 20-second window. build_quote/2 read ticker["time"] — a field that does not exist on the row. Confirmed against Coinbase's own CDP API reference for the ticker channel, independently, twice: the timestamp lives on the message envelope ("timestamp", one per frame), never on the individual tickers row. Every hand-built test fixture in this package — including the ones ported from the host adapter's own test suite (baseline_test.exs, "Phase 5.7") — encoded the identical wrong assumption, which is why this passed every test ever written against it and only ever failed against a genuine live socket. dispatch/2 now reads the envelope's own timestamp and threads it down to build_quote/3; the per-row field is gone.

  Applied the same fix to l2_data/OrderBook, which had a related but different defect: deliver_book/2 didn't read any venue timestamp — it substituted DateTime.utc_now/0 unconditionally, which is the exact substitution this file's own moduledoc already named as wrong for the ticker path (Core.Types.Quote's "never substitute now" principle) while doing it anyway one function down. deliver_book/3 now reads the same envelope timestamp and fails closed if it's absent, same as build_quote/3 — the maintained book state still updates either way, only the outgoing delivery is withheld.

  Does not, on its own, explain why level2/OrderBook delivered zero data in any of the three live tests run while chasing #22 — the old DateTime.utc_now/0 fallback always succeeded, so this was never why level2 was silent there. That remains open.

• A level2 capacity refusal from the venue was reported as :credentials_rejected — DpCryptoManagement's issue #22, filed as a suspected regression of #20. Coinbase answers both a genuine auth failure and "too many L2 streams requested in a single session" through the identical {"type":"error","message":...} frame shape. Socket.dispatch/2 collapsed both into :credentials_rejected — the shape the original stub-token incident produced — which sent a consumer that finally wired subscribe_notices/1 looking for a broken credential that was never broken. Now classified by message content: a capacity refusal reports :rate_limited, Core's own kind for pressure rather than identity: everything else keeps the original :credentials_rejected behavior.

  This does not, on its own, explain or fix why 4 of 5 shards deliver nothing. #20's fix addressed a genuine, confirmed bug (an unstaggered connect burst) but issue #22's live evidence — the refusal persisting unchanged across 15+ minutes and two clean restarts, with every socket healthy and connected — describes a permanent per-shard rejection, not the transient reset #20 targeted. Whether Coinbase enforces level2 session capacity per account rather than per connection, which would make multi-socket sharding for this channel fundamentally incompatible with this venue regardless of spacing, is not something this repository can verify without live credentials. Left open pending that evidence.

• A scope wide enough to need three or more shards opened them all in the same instant instead of staggered, and 60-second resubscribes re-issued the same burst every minute — DpCryptoManagement's issue #20, a real ~406-symbol/5-shard production scope where 4 of 5 shards (400 symbols) never delivered a single tick while the fifth did. reshard/1 scheduled every shard past the synchronous first one with the same fixed @shard_spacing_ms delay rather than one increasing per shard, so all of them opened together — exactly the connect burst this module's own moduledoc already named as the failure the venue answers with resets. Only a suite exercising three or more shards could have caught it; the existing test only ever covered two (one synchronous, one staggered), where a single fixed delay is indistinguishable from a correct one. Fixed by scheduling each shard's turn position * @shard_spacing_ms after the one before it, applied to both the initial open and the unconditional 60-second resubscribe. A regression test now exercises three shards.

• Feed.fan_out/2 crashed on a subscriber registered by name — DpCryptoManagement's issue #15. subscribe/2's to: option accepts any value, and fan_out/2 called Process.alive?/1 on it directly — which only accepts a pid and raises on anything else. A consumer registering itself under a name (ordinary OTP practice) and handing that name to to: crash-looped the whole Feed GenServer on every delivery. Fixed by resolving a subscriber (pid or name) to a pid first, treating an unregistered name the same as a dead pid: silently skipped, never a crash.

• feed_test.exs's own fake sockets never answered WebSockex.send_frame/2's internal :gen.call, silently turning several tests into a real, load-dependent race against two independent ~5-second timeouts (WebSockex's own hardcoded one and :sys.get_state/1,2's default) rather than a fast, deterministic assertion — the file's slowest tests ran 5–15 real seconds each and occasionally lost the race outright under load from the rest of the suite. Not flakiness to route around: traced to a root cause and fixed there. One fake now replies immediately per :gen's own reply protocol (removing the stall entirely); the other, which intentionally models a socket whose frames fail, now fails immediately rather than by never replying. Full feed_test.exs run time: ~45s → ~3s.

• to_order/1 read both Order.quantity and Order.filled_quantity from the same venue field — DpCryptoManagement's issue #12. order["filled_size"] populated both, so a fetched order's remaining_quantity (quantity minus filled) was always zero, even for a genuinely open, partially-filled order — a correctness bug for anything reconciling open-order state. quantity now reads the venue's own record of what was requested, from order_configuration's leaf base_size — the same field closing_configuration/1 already reads for a closing order's size, on the same response envelope. A quote-sized market order's leaf carries quote_size instead, with no rate here to convert it, so quantity is nil rather than a guess in that case.

Added

• level2 is subscribed and decoded — streamable gains :order_book. The channel was recognised and had working auth machinery since an earlier release but was never actually requested; capabilities/0 said [:quotes] while the code that would have served :order_book sat unused. Socket now maintains a real per-symbol book — snapshot then patched by update deltas, new_quantity: "0" removing a level — and delivers Core.Types.OrderBook sorted best-price-first on every change, matching this family's existing convention (see Schwab's book services) of emitting on every venue frame rather than throttling client-side.

• Sharded — this venue's whole subscription no longer runs on one socket. Measured 2026-08-27 against a live ~400-symbol universe: a level2 subscribe over the venue's real per-session limit gets "too many L2 streams requested in a single session" and the socket closes, a total data gap rather than degraded coverage — 355 of 405 pairs went stale, 1,480 refusals in one log window. Feed now opens one socket per 100 symbols (the number from that incident, carried over rather than re-derived), spaced to avoid a connect burst, level2 subscribed before ticker on each and the two spaced apart so a snapshot decode in progress does not turn a ticker subscribe into a send_timeout.

• A reconnect now resubscribes. WebSockex reconnects a dropped socket on its own and leaves it subscribed to nothing — silently, since a connected socket receiving nothing looks the same as a quiet market. Feed re-issues every shard's current subscriptions on a 60-second timer, unconditionally; the reference implementation this replaces lost a venue's entire coverage to exactly this gap for roughly forty minutes before anyone noticed the chart had gone flat.

• level2 is skipped for a credential-less subscriber rather than failing loudly for no reason. It requires a credential and ticker does not; a caller with no credentials only ever wanted the public channel, and sending a doomed authenticated subscribe would either surface credentials_required as this call's synchronous result — masking that ticker works fine — or cost a wire round trip to learn what the credential's absence already answers.

Documentation

• The :unsupported list is now split. venue_does_not_serve/0 names the 38 endpoints that are Coinbase's own absence — staking reads, the one-step convert, funding rails, option chains, watchlists — each with the source and date behind it; three (get_funding/2, get_contract_stats/2, list_instruments/1) stay under @not_ported because they are the venue's surface and this package's backlog, not the venue's gap. Robinhood found four callbacks mislabelled the other way; this pass checks Coinbase's own list rather than assume it was filed correctly the first time.
• README.md states what the contract covers — 46 of 87 callbacks :experimental, and points at negative-claims.md for every absence's source.
• docs/reference/coinbase/endpoint-inventory.md's counts refreshed. It read "everything authenticated is absent" until this release, which had been true at capture and stopped being true as this package grew — the vendor-side numbers had not moved, this package's coverage of them had, and the section conflated the two.

Documentation

• Every negative this package makes is audited — docs/reference/coinbase/negative-claims.md, twelve claims with the source and date consulted for each. Nine hold; three were wrong, and all three for the same reason: each was a true statement about one endpoint restated as a claim about the venue.

  supports_order_preview: false and supports_order_replace: false were assumed without reading the list the endpoints are on — the second mattered more, because it told a caller to cancel and re-place, opening a window in which no order is live. And get_trade_volume/2's "Advanced Trade does not aggregate" was read off /products/volume-summary, which is market volume and a different question.

  The check that would have caught all three is the one the table now enforces: name the endpoint you looked at, and the date.

• usage-rules.md gains the surface this release added — the two accounts a futures position is margined from, Prime's separate host and credential triple, convert's absent expiry, portfolios as addresses, and the fee/volume pair.

• AGENTS.md gains a pointer to this package's own usage-rules.md, so a reader who opens the generated file knows where the package's rules actually are.

Changed

• Core dependency moves to ~> 0.1.36, and place_orders/3 is declared absent with the reason: this venue places one order per request. A batch is one request the venue accepts or rejects as a unit, and a caller placing several here calls place_order/3 several times and reconciles the outcomes itself.

Added

• Key permissions and the server clock — get_roles/1, get_server_time/1 and a test_connection/2 that is no longer declared absent.

  can_transfer is a separate permission from can_trade, and a key routinely holds one and not the other. Asking is cheaper than discovering a missing one from a refused withdrawal. The response also names the portfolio the key is scoped to, which is where a caller finds out whose balance it has been reading.

  test_connection/2 asks two different questions and picks by what it was given. Without credentials it reads the public clock — reachability alone. With them it reads the key's permissions, which fails if the key is wrong and answers what the key can do if it is right. An unreachable venue and an unaccepted key are different problems.

  get_server_time/1 returns the venue's own map undiffed. The difference a caller cares about is against its own clock at the moment it asked, and computing it inside the package would hide the round trip in the number. It is worth reading at all because this venue's JWT window is two minutes: a host clock further out than that produces authentication failures that look like a credential problem.

• Convert, portfolios and the transaction summary — the last ten Advanced Trade endpoints in the coverage plan's Phase 11.

  Convert is the facade's only two-step operation, and Advanced Trade states no expiry at all. expires_at is nil, which means "not stated" and never "open-ended": a caller committing a lapsed quote can be filled at the current rate rather than refused, which is the dangerous outcome because the operation looks like it succeeded and every number is real. commit_conversion/2 and even get_conversion/2 re-ask for both accounts — the venue's own rule, unusual for a read — and this package fills neither in: a conversion committed against accounts the caller did not name happens between the wrong two balances. A status this package does not know maps to nil, never the nearest one.

  A portfolio is an address, not a value. list_portfolios/1 returns them, get_portfolio_breakdown/3 returns what is inside one — a different and much larger answer — and create_account/1 and rename_account/3 reach the portfolio endpoints, because Advanced Trade has no notion of creating an account. Deleted portfolios stay in the listing: the venue keeps them because old orders still name their ids, and filtering them out would make a historical id look like one that never existed.

  get_trade_volume/2 was declared absent on a claim that was wrong. This package held that "Advanced Trade does not aggregate" the account's own volume; the transaction summary does, in volume_breakdown per volume type with advanced_trade_only_volume and coinbase_pro_volume beside it. The claim had been made from the market volume endpoint's absence, which answers a different question. The two account totals ride alongside the breakdown rather than being folded in: the venue documents the first as non-inclusive of the second, so adding either to the breakdown double counts.

  get_fees/2 carries both fee_tier and fee_tier_without_promotion — they differ while a promotion is running, and it can end between two calls — and keeps the tax's INCLUSIVE/EXCLUSIVE flag, because the same rate quoted either way is a different amount of money.

• US derivatives — the nine CFM endpoints. get_positions/1 and list_futures_positions/1, get_futures_position/3, get_futures_balance_summary/2, the three sweep calls, and the three intraday-margin calls.

  Two accounts, and the balance summary names both. Futures margin from an account held with Coinbase Financial Markets; spot sits in one held with Coinbase Inc. cfm_usd_balance is the first, cbi_usd_balance the second, total_usd_balance the pair — and a caller sizing a futures position against the total is sizing against money that is not there. Every amount keeps its currency; flattening it off is how two currencies get added.

  :realised_pnl is nil on a Types.Position from this venue, and that is not an omission. Coinbase publishes daily_realized_pnl — what the position realised today — and no lifetime figure. Putting a daily number in a field that means the position's answers a different question under the same name: a caller summing it across reads counts one day repeatedly. The daily figure is not discarded — list_futures_positions/1 returns the venue's own row, where it keeps its own name, along with expiration_time, which Types.Position has no place for either because a future expires and a perpetual does not.

  A sweep is scheduled, not settled. schedule_futures_sweep/2 queues a move out of the futures account and list_futures_sweeps/2 reports the queue; a listed sweep has not happened. Omitting the amount sweeps every available excess dollar — the venue's documented default, stated here because a caller reading a missing amount as "nothing" would move the lot. cancel_futures_sweep/2 cancels the pending sweep and takes no id.

  INTRADAY_MARGIN_SETTING_UNSPECIFIED is not _STANDARD. It is the venue declining to say, and mapping it to the safer-sounding value would assert a setting the account may not have. The venue's own strings are returned and required on the way in, with no default: UNSPECIFIED is a value in the enum, and choosing it for a caller would set the account to something it did not ask for.

  get_current_margin_window/2 carries both kill-switch flags. An account that believes it is on intraday margin while the switch is enabled has more leverage in its plan than in its account.

  supported_instrument_types gains :future. :perp stays absent: Advanced Trade's perpetuals are the INTX endpoints, which are APPROVED-SKIP as deprecated, and declaring a surface this package does not reach would be a claim about the venue standing in for one about the package.

• Coinbase Prime custodial staking — DpExchange.Coinbase.Prime, all nine endpoints, with stake/3 and unstake/3 now live on the facade.

  A different product, host and signing scheme. Everything else in this package talks to api.coinbase.com/api/v3/brokerage and signs a CDP JWT; Prime talks to api.prime.coinbase.com/v1 and signs an HMAC under an access key, a passphrase and a signing key that Advanced Trade neither issues nor accepts. Two of the three credentials is {:error, :missing_prime_credentials} rather than a request that is signed and wrong.

  These are not the CDP Staking API. Those seven are on-chain: they take a wallet address and return unsigned transactions for the caller to sign and broadcast. Reaching them through stake/3 would be this family's recurring failure at its most expensive — a caller believing it had staked while holding a transaction nobody sent.

  Two scopes, and this package picks neither for you. Prime publishes every staking operation across a portfolio and again on one wallet, and the two are not interchangeable: a portfolio-scoped unstake redeems across every wallet in the portfolio. stake/3 and unstake/3 follow only what the caller said — a :wallet_id means the wallet, its absence means the portfolio — and opts[:portfolio_id] is required, refused as {:error, :missing_portfolio} before a request is made.

  Four callbacks stay declared absent with the reason: Prime publishes no rate schedule and no staking history at either scope; staking/status names one wallet and is not "every staked position, one per asset" (reachable as Prime.staking_status/4); and claim_rewards is a write that moves accrued rewards, not a report of what accrued.

  Nothing here has been run against Prime. The paths are read from the vendor's pages on 2026-08-31 — thirteen pages, nine endpoints, four pairs documenting one path under two names — and the signing scheme from Prime's authentication documentation. This repository holds no Prime credential and money-moving endpoints are answered in production, not by a test here. Responses come back as the venue's own maps for the same reason: a Types.StakingBalance built from an unverified field name is a plausible number in the wrong field.

• Payment methods and the internal move: list_payment_methods/2, get_payment_method/3 (GET /payment_methods, GET /payment_methods/{id}) and transfer_internal/4 (POST /portfolios/move_funds).

  A payment method's flags disagree with each other. Each row carries verified, allow_deposit and allow_withdraw, and a method verified for deposit is routinely not verified for withdrawal. Rows stay the venue's own maps and no "usable" boolean is synthesised from them — collapsing the flags is what makes a caller move fiat through a method the venue refuses.

  get_payment_method/3 is the read; the listing is a snapshot. A method's state changes without the account doing anything, and selecting the row out of an earlier listing answers with whatever was true when that listing was taken.

  transfer_internal/4 moves nothing off Coinbase — no chain, no address, no network fee. Both portfolio uuids are required and neither is defaulted: a move missing either is {:error, :missing_portfolio} before a request is made, because the alternative is shifting funds between portfolios the caller never named. The amount is sent in full notation, since Decimal.to_string/1's scientific form is not a number this venue reads.

Changed

• Core dependency moves to ~> 0.1.33, and with it twelve callbacks are now declared rather than missing. Nine are declared absent with the reason, checked against the venue's own reference on 2026-09-01: Advanced Trade publishes no allowlist (request_approved_address/4, remove_approved_address/3), no networks list (list_networks/2), no fiat registration (add_payment_method/2), no fee promotions (list_fee_promos/1), no FX publication (get_fx_rate/3), no notional valuation (get_notional_balances/3) and no custody product (list_custody_fees/2).

  get_transactions/2 is absent for a different reason worth stating. /transaction_summary exists and is not it: that endpoint reports what the account traded in a window and what it cost, not an enumeration of deposits, fees and adjustments. Returning it here would have answered a different question while looking like this one.

• quantization/1 — what the venue will actually accept, and Rest.get_product/2 for the whole record. Both were :unsupported.

  The venue names four increments and they are not interchangeable. quote_increment bounds the price and base_increment the quantity; a caller rounding a price to the base increment produces an order the venue rejects on a field it did not name. Both minima are carried too — base_min_size is units and quote_min_size is cash, and a market order sized in cash is bounded by the second where a limit order in units is bounded by the first.

  status is the venue's own word, unmapped: a boolean would lose the difference between a product that is paused and one that is gone.

• get_symbols/1 reads the authenticated catalogue when a credential is present. Third and last of the public/private path corrections — the book, the candles and now the product list were all reading /market/… regardless.

• get_trades/2 — the public tape. get_price/2 already reads this payload and keeps only the newest print, because a Quote has room for one price; the rest were discarded at the boundary. This returns them.

  Not get_trade_history/2, which is the credential's own fills. broken is false on every print — the ticker publishes no bust flag, and a venue with nothing busted reports nothing busted.

• get_historical_prices/4 reads the authenticated candles path when a credential is present. The venue publishes the same candles twice — /market/products/… public and /products/… for a credential — and this always called the public one, so a caller holding a credential was silently forgoing whatever the authenticated view adds. Same correction as the product book.

• get_order_book/2 — depth, which this package declared :unsupported. GET /product_book for a credential and /market/product_book without one — the venue publishes the same book twice, and reading the public one while holding a credential would silently forgo whatever the authenticated view adds. The venue's limit and aggregation_price_increment are passed through.

  Both sides come back as the venue ordered them. Re-sorting would hide a venue that sent a crossed or out-of-order book, which is exactly the thing worth seeing.

  A book the venue did not date is refused. A depth snapshot carrying the client's clock cannot be told apart from a current one, and a stale book read as current is the most expensive wrong number here. sequence stays nil — the endpoint publishes none, and a caller must not learn to detect stream gaps from a REST book.

Fixed

• get_top_of_book/2 now carries the sizes. It read /products/{id}/ticker, which publishes best_bid and best_ask and nothing about how much is there — so bid_size and ask_size were nil on every response.

  That nil was honest and it was avoidable: the venue publishes /best_bid_ask, whose pricebook carries the size at each level. A price without a size is half a top of book — a caller sizing against the best bid needs to know whether there is 0.01 there or 40, and nil gave it no way to ask.

  An empty side is still nil rather than zero: one side of a book can genuinely be empty, and zero would claim someone is quoting nothing at a price of nothing.

• get_trade_history/2 — past fills.

  trade_type is not decoration. Regular fills carry FILL; the venue also emits REVERSAL, CORRECTION and SYNTHETIC for adjusted ones, and a reversal is not a trade that happened. Core.Types.Fill has no field to say which is which, so summing a mixed list produces a position and a cost basis that are both wrong and both plausible. This returns only FILL rows by default, and opts[:trade_types] widens it — returning all four under a type that cannot distinguish them would be a substitution, and refusing them entirely would hide corrections the venue made.

  A fill the venue did not date is refused, not stamped with the local clock: a fill is an event at a moment, and a client timestamp places it wrongly in a history while looking entirely reasonable.

  fee_currency is nil rather than the pair's quote guessed from the symbol — a fee can be charged in a third asset and often is. UNKNOWN_LIQUIDITY_INDICATOR maps to nil, because neither :maker nor :taker is an honest answer to the venue saying it does not know.

  Filters go to the venue rather than being applied to the page it returned, and the walk follows cursor to a page bound.

• get_balances/2 and get_accounts/2. The package could not say what the credential holds.

  The venue reports available_balance and hold and no total. The total here is their sum — arithmetic on two numbers the venue stated, not an estimate — and it is nil when either is missing rather than the other one alone. "Available 1.25, total unknown" and "total equals available" are different claims, and a consumer sizing against the second when the first is true trades against money that is held.

  The endpoint pages, at 49 by default and 250 at most, and this follows the cursor. A caller reading one page holds some of its balances with nothing to say which are missing, and every number on that page is real — which is what makes stopping there worse than failing. @max_account_pages bounds it, so a server that always says has_next errors rather than looping inside a facade call.

  get_accounts/2 is separate because an account is more than a number: a caller routing an order needs the uuid and the platform, and a caller sizing one needs the balance. Collapsing them would lose the first. opts[:uuid] reads the single-account endpoint.

  :timestamp is when the request was made — a balance has no venue event time.

• convert/4 and get_trade_volume/2 (Core 0.1.22) are declared unsupported, with the reasons checked. Advanced Trade's convert is the two-step form — POST /convert/quote, POST /convert/trade/{id}, GET /convert/trade/{id} — which is quote_conversion/4 and friends, scheduled separately. The one-step POST /conversions belongs to the Exchange API, a different product this package does not reach. /products/volume-summary is market volume and lives there too; get_trade_volume/2 asks what this account traded, which Advanced Trade does not aggregate.

• preview_replace/4 and close_position/3. Two documented endpoints this package had no facade for.

  POST /orders/edit_preview prices an amendment before it is made. It is not preview_order/3 with an order id: the venue prices the amendment against the resting order's own state, including whatever of it has already filled, and its response carries average_filled_price and order_margin_total — numbers a fresh order does not have. It takes the same :price / :quantity change set replace_order/4 does and refuses anything else before the request.

  POST /orders/close_position flattens a position by having the venue place the closing order. The returned Order carries no side. The venue never states one, and it worked the side out from a position this package did not read — filling in :sell because closing is usually selling is wrong exactly where it matters, on a short. The order type, time in force and size are read, from the order_configuration the venue echoes back, and a configuration key this package does not recognise leaves them nil rather than picking the nearest.

Fixed

• cancel_all_orders/2 is declared unsupported, with the reason checked. POST /orders/batch_cancel takes an explicit order_ids list — it is the endpoint cancel_order/3 already uses, one id at a time. There is no "cancel everything" call here, and assembling one from get_orders/2 plus a batch would be N partial outcomes with no way to reach an order that appeared between the listing and the cancel.

• BREAKING: get_historical_prices/4 returns Core.Types.Candle. It was returning Quotes with price: close.

  The venue sends open, high, low and close for every bar. Three of them were discarded here, at the boundary, where no caller could see it happen — and everything that came out was a real number, so nothing looked wrong. A caller reading price was holding one corner of a bar with no way to learn it.

  This is the same defect the coverage plan's 2.10 found in Schwab, with the same reasoning behind it, still live here after that one was fixed. The fake had it too: it returned get_price/2's Quote, so the suite agreed with the bug it existed to catch.

  Bars now carry all four prices and :opened_at — the venue's own bucket start, used as-is. A bar the venue did not date is refused with :missing_venue_timestamp rather than stamped with the local clock, which would place it wrongly while looking right.

Fixed

• This package claimed the venue has no order preview and no atomic replace. It has both. supports_order_preview and supports_order_replace were declared false on those claims, and neither was checked against the venue's reference. Coinbase publishes POST /orders/preview and POST /orders/edit; both flags are now true and both endpoints are implemented.

  The replace claim was the worse of the two. Its moduledoc called supports_order_replace: false "a claim about risk rather than convenience", because cancel-then-replace opens a window in which no order is live. The risk was real and the claim was wrong: the package was describing a hazard it was creating by not implementing the endpoint that avoids it.

Added

• preview_order/3 builds the same order_configuration as place_order/3, so a preview is a preview of the order that would actually be sent. A 200 carrying a populated errs is a refusal — returning it as a successful preview would tell a caller its order is fine when the venue has already said otherwise. A warning is passed through and does not make it a refusal.
• replace_order/4 edits price or size in place. Any other change is refused rather than dropped: a caller trying to change the side is describing a different order, and editing only the price would leave it holding one it did not ask for. The venue's edit response carries no order body, so the order is read back rather than reconstructed from the request — reporting what was asked for as though the venue had confirmed it is the mistake this whole contract is written against.

Added

• cancel_order/3, get_order/3, get_orders/2. The order lifecycle, where there was none.

  Cancellation is a batch endpoint that refuses per order. POST /orders/batch_cancel answers with a results array carrying its own success and failure_reason per id, so a 200 says nothing about whether anything was cancelled. A batch of one is still a batch. An order already filled comes back as a refusal, not an :ok — "I cancelled it" and "it was not there to cancel" are different facts, and a caller retrying on the second is chasing nothing.

  CANCEL_QUEUED maps to :open, not :cancelled. An order accepted for cancellation is still live until the venue says otherwise; reporting it gone invites a second order for the same exposure.

  A status, side, order type or time-in-force this package does not recognise is nil, never the nearest atom. A venue adding a word later produces an absent field rather than a plausible wrong one.

  get_orders/2 filters at the venue rather than in this package — a client-side filter over one page would silently drop matching orders sitting on the next. It returns one page and does not follow the cursor, which is stated rather than left for a caller to discover while reconciling.

• place_order/3. This venue could not place an order; it can now.

  Coinbase names the order type and the time-in-force in a single key — limit_limit_gtc, market_market_ioc, stop_limit_stop_limit_gtd — and the set of names is sparse. There is no limit_limit_ioc, no market_market_gtc.

  A pair the venue does not name is refused before the request is sent. Sending {:limit, :ioc} as limit_limit_fok would place an order that fills-or-kills where the caller asked for immediate-or-cancel, and every field in the request would look right.

  Three further refusals rather than defaults: a limit without a price, a stop-limit without a stop price, and a market order sized in neither base nor quote. post_only is omitted when unset rather than sent as false, because silence is not a decision to take liquidity.

  A 200 carrying success: false is a refusal, not a placed order.

  client_order_id is the venue's idempotency key: a caller's own is passed through, and a v4 UUID is generated from the VM's CSPRNG when absent.

Added

• DeprecatedEndpointsTest — fails the build if any code path constructs one of Coinbase's six vendor-deprecated INTX endpoints. They are absent today; nothing kept them absent.
• docs/reference/coinbase/endpoints-enumerated.tsv and a rewritten inventory: the documented surface is 712 REST operations and 46 socket channels, enumerated endpoint by endpoint from all 806 reference pages, replacing a page count. Deribit alone was recorded as 37 and is 115 — Coinbase renders it as twelve sibling trees with no deribit in their paths.
• Prime's custodial staking enumerated: 13 documentation pages, 9 endpoints, four pairs being duplicate pages for one path.

Added

• Repo scaffold from the DpExchange standard; extraction pinned to the host's 553fa787 with its working-tree state recorded, since the Coinbase subtree was dirty at extraction time.