# TDD log GSM 03.38 was developed RED/GREEN. 1. RED: 7 contract tests, 7 failures. The missing API, table inventory, independent pairs, malformed-input behavior, 39 names, and registration were all specified before implementation. 2. GREEN: the initial implementation passed all 7 contract tests. 3. RED: `valid_pairs/0` was specified; 14 tests ran with the new contract as the sole failure. 4. GREEN: all 14 tests passed with the complete 182-pair inventory. 5. Conformance GREEN: exhaustive cell tests, ICU differential fixture, wrapper wiring, lifecycle, UTF-8 errors, and linear discard coverage brought the suite to 17 tests with 0 failures. 6. Performance refactors were accepted only while the full suite remained GREEN; the benchmark retains the displaced implementations as A/B baselines. 7. RED/GREEN cycles added failing vectors first for SMS septet packing, TBCD, and all three UCS2 alpha-identifier forms. That suite had 29 tests. 8. RED: six ITA2 tests failed because the codec, state machine, packing API, exhaustive tables, and aliases did not exist. GREEN: all six passed; the package contract then intentionally remained RED until ITA2 was added to the support and conformance records. 9. RED: six CCIR 476 tests and the package contract failed before the 32 traffic signals, service signals, seven-bit packing, FEC inversion, and SITOR/NAVTEX registration existed. 10. GREEN: implemented every M.476-5 table value, classified all 128 possible input units, and reused the exhaustively tested ITA2 state machine. 11. RED: six AIS6 tests and the package contract failed before Table 45, six-bit packing, IEC 61162 armoring/fill-bit validation, and aliases existed. 12. GREEN: implemented and exhaustively tested all 64 Table 45 values, strict malformed input, exact field packing, both armor ranges, and zero-fill validation; the focused six-test suite passed before documentation was updated to restore the full package contract. 13. RED: six ITU-T S.2 tests failed before its two-mode case state machine, single-capital behavior, three-capital lock optimization, initialization, and aliases existed. 14. GREEN: implemented the official S.2 state diagram and passed isolated, locked, figure-transition, initialization, full-alphabet, malformed-input, and registration vectors. 15. RED: the package contract had two failures when TBCD and SIM/USIM alpha identifiers were required as registered Iconvex codecs rather than helper APIs only. 16. GREEN: added strict codec adapters, direct UTF-8 paths, discard behavior, aliases, lifecycle registration, and end-to-end `Iconvex.convert/4` vectors; the focused helper/contract suite passed 12 tests before documentation restored the package-wide support assertion. 17. RED/GREEN: ITA1 began as six failing contracts for its complete 32-row polarity table, two state tables, national/non-text holes, all-octet classification, aliases, and five-bit packing. The complete table port made those six tests GREEN. 18. RED/GREEN: US-TTY and MTK-2 each began with six missing-codec failures. GREEN required every source-table cell, every register transition, complete repertoire round trips, high-bit rejection/discard, aliases, and pinned source identities. 19. RED/GREEN: ITA3 and ITA4 began with five contracts apiece. GREEN required all 32 traffic conversions, classification of all 256 octets, preservation of the ITA2 state machine, exclusion of service signals, and aliases. 20. RED/GREEN: T.50/IA5 began with six contracts for all 128 positions, the complete invalid high-bit domain, seven-bit packing, names, source pin, and UTF-8 boundaries. The exact IRV implementation made all six GREEN. 21. RED/GREEN: International Morse began with seven contracts for all 51 graphic assignments, six procedural signals, explicit serialization, token offsets, discard/noninjective behavior, aliases, and the official source pin. The completed M.1677 implementation made all seven GREEN. 22. RED: the strengthened package contract exposed that seven GREEN codecs were missing from the published 45-codec list; a second failure required a nonexistent generated runtime inventory. 23. GREEN: the package now generates a 52-row canonical/alias/module/stateful inventory directly from runtime modules. Tests compare every field and the documentation names every registered canonical codec, preventing this class of drift from recurring. 24. Performance RED/GREEN: a retained map-lookup baseline showed the generic US-TTY/MTK-2 state engine lagging the specialized ITA2 loop. The optimized loop holds active decode/encode tables across ordinary characters and uses a compile-time cross-register index; all 12 focused contracts remained GREEN. 25. Alternating 12-run medians measured +46.3%/+20.1% US-TTY encode/decode and +34.3%/+14.7% MTK-2 encode/decode. The displaced implementation remains in `bench/five_unit_shift_benchmark.exs` as the reproducible A/B oracle. 26. RED: four packed-inventory contracts failed before a discoverable unified facade existed. GREEN: 49 profiles cover every fixed-width registered codec, both bit orders, exact standard-order metadata, the GSM published vector, runtime exclusions, and a generated CSV; 4 tests pass. 27. Performance RED: profile lookup rebuilt and sorted all 49 metadata maps on every message (~159 ms per 10,000 lookups). GREEN: immutable profiles and name/module indexes are cached in `persistent_term`; the focused contracts remain GREEN. 28. A/B GREEN: on 65,540 septets, the bounded streaming implementation measured 173.5× faster packing and 59.6× faster unpacking than the retained whole-message GSM integer implementation. The benchmark executes both paths and verifies every returned result. 29. Review RED/GREEN: 16 focused tests initially produced 9 failures covering malformed SIM/TBCD fast callbacks, compressed-surrogate safety, strict GSM option validation, octet-atomic TBCD discard, UCS2 padding boundaries, ICU-pinned ESC fallbacks, packed-order errors, exhaustive extension cells, and application registration ownership. The focused suite then passed 16/16 and the full package passed 108/108. 30. Registry-recovery RED/GREEN: the pre-fix worker-crash reproduction lost registrations while `:iconvex_telecom` remained started. The inherited registry now preserves GSM0338, TBCD, every other Telecom entry, and their exact application-owned tokens. The regression stops Telecom after recovery and proves token cleanup removes both named codecs; the complete focused review-regression file passes 13/13. 31. Late-startup rollback RED/GREEN: a deterministic conflict on the final SIM-alpha codec was specified before changing production. The first focused run was RED (14 tests, 1 failure) because the assertion omitted OTP's application-start error envelope. GREEN normalizes that envelope and proves all 51 earlier codec modules and every one of their names are removed, the preexisting conflicting registration and ownership token survive, and a clean 52-codec restart succeeds. The existing ownership rollback needed no production change; the focused file passes 14/14. 32. Consumer-artifact RED/GREEN: the release audit and a new package contract reproduced that `test`, `bench`, and `tools` were selected for Hex. The corrected manifest keeps runtime assets, source/license notices, generated inventories, conformance evidence, and public documentation while leaving development corpora in the repository. 33. Managed-set RED/GREEN: cross-package startup first failed because Specs and Extras had retained overlapping names. Telecom now publishes all 52 codecs in one atomic managed set. A registry-repair RED also caught a transient missing `GSM0338`; repair now republishes complete name indexes before pruning stale rows. The full-stack harness passes all six extension start orders with 1,990 unique canonical names, and Telecom's complete suite passes 115/115 with warnings as errors. 34. Release-cardinality documentation RED/GREEN: the executable contract was added before changing prose and failed 1/7 on the stale 1,995 full-stack count. Current release prose now states the derived 2,005 cardinality; the package contract passes 7/7 and the complete suite passes 116/116 with warnings as errors. 35. Telecom-source provenance RED/GREEN: the two-contract suite was written before `SOURCE_PROVENANCE.md` or its package selector and failed 2/2 (`seed: 0`). GREEN inventories every retained `tmp/` artifact by exact SHA-256, official parent or retrieval URL, ownership, and derived-file relationship. One 245-byte rejected-request HTML response masquerading as `itu-r-m1677-1.pdf` was removed. The retained S.13 PDF is correctly identified as the 1976 CCITT Orange Book despite its historical local filename. All upstream evidence is repository-only, no license is invented, and Hex packaging excludes `tmp/`. The focused suite passes 2/2 in 0.1 s; the complete package passes 118/118 with warnings as errors. The rebuilt Hex archive was inspected and contains no `tmp/` member. 36. IBM Transcode RED: two new test files were committed before runtime code, source assets, inventories, or the benchmark. The focused run produced 14 tests and 14 failures: both modules and registrations were absent, both normalized source tables and metadata were absent, no packed profiles or inventory rows existed, and the benchmark script did not exist. 37. IBM Transcode GREEN: separate GA27-3005-3 and GA27-3004-2 codecs implement all 64 cells while preserving their sole `0x0C` disagreement. The suite classifies all 256 octets for both profiles, checks every Unicode scalar, exercises strict/discard/substitute/direct UTF-8 paths, verifies every source and UTF-8 stream split plus absolute error offsets, and pins complete MSB/LSB vectors, all packed tails, invalid lengths, nonzero padding, and transport types. The focused suite passes 14/14. 38. Source/package GREEN: both primary PDF URLs, sizes, SHA-256 values, exact table pages, normalized 64-row CSV digests, and low-order-first semantics are pinned. Only the original CSV facts and metadata are packaged; raw IBM PDFs remain excluded. Runtime inventories now contain 54 codecs and 51 packed profiles (4 six-bit). 39. Performance GREEN: two production runs measured 14.792–49.251 MiB/s over twelve codec/transport paths. Native/reference latency was 0.704×–0.890×, and all four deterministic reduction-scaling gates were 1.881×–1.890×. The executable benchmark enforces the 30× ceiling and linear scaling. 40. Packed-name/benchmark review RED: suffix-resolution, transport-tag, benchmark, and run-provenance contracts were added before fixes. The focused 8-test run failed exactly 4 tests: packed names were unresolved, all mistagged LSB containers decoded, packed benchmark rows had no comparison, and the record contradicted itself about two versus three runs. 41. Packed-name/benchmark GREEN: all 102 canonical packed names and 220 packed alias forms now select their named order; explicit conflicts and all 51 mistagged LSB profiles return `:bit_order_mismatch`. Independent bit-buffer references call no native packed API, compare exact output for both profiles and all six operations, enforce 12 relative ceilings, and enforce 12 linear reduction gates. Two production runs measured 13.934–42.258 MiB/s, 0.439×–0.891× native/reference latency, and 1.685×–1.979× scaling. The focused suite passes 8/8, the full package passes 135/135 with warnings as errors, and a forced production compile passes with warnings as errors. 42. ADVERSARIAL PACKED/REFERENCE RED/GREEN: four semantic packed-decode vectors were added first at a nonzero unit index for ITA1, ITA3, ITA4, and CCIR476 in both bit orders. RED returned unpacked byte offset 2 and an octet fragment; GREEN translates it to the physical `unit_index * unit_bits` offset and preserves exact-width MSB or integer LSB fragments. A separate benchmark-source RED rejected the self-derived `codec.table()` oracle. GREEN digest-validates and parses both packaged IBM 64-row CSVs, builds the decode/encode reference independently, and retains the independent packed bit-buffer paths. The focused suite passes 9/9. 43. Full-stack release-cardinality RED/GREEN: the executable package contract was changed first from 2,064 to the regenerated 2,100-codec registry and failed 1/7 against both stale release documents. README and changelog now state 2,100; the focused contract passes 7/7. 44. GSM TP-UDHL DOMAIN RED/GREEN: an adversarial boundary regression first failed 1/1 because the UDH alignment helper accepted 256 even though TP-UDHL is one octet. GREEN limits the documented TP-UDHL input to 0..255; 255 remains aligned correctly while negative, 256, and noninteger values return the existing typed `:invalid_udh_octets` error. The focused regression passes 1/1 with warnings as errors. 45. REQUIRED EXTERNAL SUBSTITUTION CALLBACK RECURRENCE RED/GREEN: after Core made the linear `encode_substitute/2` path an explicit external-codec contract, the complete Telecom RED failed 1/137 because its late-startup conflict fixture still implemented the former callback set. GREEN gives that test-only codec the required callback without changing the conflict scenario. The focused lifecycle run passes 2/2 and the complete package passes 137/137 with warnings as errors. 46. PROVENANCE-QUARANTINE RELEASE CARDINALITY RED/GREEN: the executable release contract changed first and failed exactly 1/7 while README and changelog still claimed the former 2,100-codec registry. GREEN records the current 2,093 unique canonical names after seven Specs codecs were quarantined; the focused package contract passes 7/7 with warnings as errors. 47. REDUCTION-HARNESS GC ISOLATION RED/GREEN: the final OTP 27 package run failed 1/137 when one IBM 2780 decode comparison charged an ordinary-heap garbage collection to only the 20,000-unit side and reported 1.579x. Independent measurement proved the same 20,000/40,000 path costs 25,438/50,601 reductions, or 2.002x after fixed overhead, in a large-heap worker. An expectation-first source contract then failed until every reduction measurement ran in a monitored fresh one-million-word heap. GREEN preserves the two-sided 1.60x..2.60x algorithmic gate, output parity, and 30x native/reference ceiling. The focused benchmark test passes 1/1; two complete local runs report 1.982x..2.050x across all twelve paths. 48. SIM ALPHA FRAMING RECOVERY RED/GREEN: the exact `80 D8 00 00 41` reproducer first failed native discard by returning no suffix, while every Stream split was rejected as unsupported; the captured 2/2 RED log SHA-256 is `8344c8dc0b2639bf2e9f7fffcf9b098e469bacf8f7b7468b09f98727d26ab7ce`. GREEN implements a bounded native state machine for GSM, `0x80`, `0x81`, and `0x82` fields. Recovery consumes the complete invalid UCS-2 unit, retains the selected frame, advances compressed payload counters, preserves trailing GSM record padding, and buffers target encoding until the final chunk. Discard keeps `A`; per-byte substitution yields `<00>A`; callback replacement yields `?A` at offset 1 for every split. A supplemental `0x82` surrogate and all four valid forms prove state and split invariance. The focused suite passes 4/4, the SIM/review/package selection passes 30/30, and the complete package passes 141/141 with warnings as errors; full-run log SHA-256 is `f3579438107eb74c54dd4775e48c8b23ead28eecea3308cb30034f7c9b516e60`. 49. SIM FIRST-ERROR ORDER RED/GREEN: `80 D8 00 00` contains an invalid surrogate unit before its odd terminal byte. Public one-shot decoding first returned only the later truncation, while every Stream split correctly returned the surrogate at byte offset 1. The expectation-first focused run failed 2/7; its RED log SHA-256 is `5d31c74403f9f81a04d436e7ad4fb3aabc3600c2b60cd46f13c9e1440b54aa11`. GREEN scans complete UCS-2 units before classifying a trailing single byte, so the low-level API, codec fast path, one-shot conversion, and all five splits return the same invalid sequence. The focused run passes 7/7; its GREEN log SHA-256 is `d6d97d5714a0cf310aab1834496a641d14d87c4bd0f9cae6455100754d89bf97`. 50. SIM DECLARED-PAYLOAD EOF RED/GREEN: counted `0x81` and `0x82` records with one valid `A` byte and one declared-but-absent terminal payload byte first retained `A` in native discard but raised at every Stream split. A separate strict-parity RED showed one-shot errors at offset 0 with the full record while incremental decoding reported physical EOF with an empty sequence. The discard and strict RED log SHA-256 values are `29ee9c545777f250f89b31221bb8f2237f788b554d66bb83696347f97abe6b86` and `dd051bbd7a79366f83bb06e290b1397ef958d280e687635cd69a76d8eaa958f7`. GREEN retains the valid prefix only for plain discard and normalizes both complete compressed headers to the exact physical EOF coordinate. Strict, substitution, and callback policies remain incomplete errors with no callback for an absent byte. The focused run passes 9/9; its GREEN log SHA-256 is `8ad404dd60a5429b9ac7679b602ee4f86a7b162b68df5d5f1a0601f4878f6eb0`. 51. SIM COMPRESSED-PREFIX PARITY RED/GREEN: a bounded matrix first compared all 25 unique proper prefixes of six representative `0x81`/`0x82` records under strict, discard, byte-substitution, and callback policies at every byte split. It includes every partial header, valid payloads, GSM extensions, surrogate and overflow bases, exact errors, outputs, and callback events. RED failed 2/2 because `82 02 D8 00 80` reported the later declared-payload EOF one-shot but the earlier invalid U+D800 unit in Stream. The RED log SHA-256 is `8d53977c59373d0b09532f7398dc1f449088620d8830d155f522f6c06ef3bf8f`. GREEN decodes every physically present compressed payload byte before classifying a later declared EOF. A trailing GSM escape with more declared payload is the physical incomplete frame rather than a local-offset leak. The final suite pins the five partial-header frames explicitly and passes 4/4; its GREEN log SHA-256 is `938dcd11c5e826a8f0a5a7938674da317a4eae1219547fa23b97c62dce2f5009`. 52. SIM COMPLETE-RECORD BOUNDS RED/GREEN: expectation-first tests exposed the unbounded target list, unrestricted source length, and missing public record limit. The clean RED failed 6/8; its log SHA-256 is `bf39de1d6d39d19cefb05b528c56a1c5e40c7f7252eb9ce085c0a017029dc58d`. GREEN enforces the ETSI TS 102 221 section 8.2 255-octet linear-fixed-record boundary in the low-level helper, one-shot codec, and Stream paths. Decoder state records consumed octets separately from pending bytes and proves `total + pending <= 255` at every split of maximum GSM, `0x80`, `0x81`, and `0x82` records. Strict overflow reports byte 255; discard, substitution, and callback recovery remain split-invariant at stable offsets 255 and 256. Earlier GSM, surrogate, compressed trailing-escape, and intrinsic target errors retain priority. Target state and every temporary admission candidate are capped at 255 code points. Exact Stream admission has a fixed worst-case quadratic cost inside that protocol cap; one-shot strict encode uses a bounded prefix and binary search. A production micro-run measured a maximum 255-character strict encode at 375.92 microseconds and 3,267.52 reductions, while the retained 255-code-point target state occupied 510 flat words. The final focused suite passes 9/9 with log SHA-256 `90f393f0320eae324189907b7c3836e4a64d4dde52ae50bb2c438bba5bebfd44`; the complete package passes 159/159 with warnings as errors and log SHA-256 `59ab5269a981ebf18836ea66f151aa1507fcc72cf28627ea0acc873ee8e5ca67`. A forced production compile with warnings as errors passes with log SHA-256 `4a38a51f5805f9738c367388016a161bee23e67c8781ce932db416e4fe79b368`. 53. SIM INITIAL-GSM RECOVERY COMMIT RED/GREEN: an adversarial audit supplied `83 80 00 41`. Strict decoding correctly rejected byte 0, but discard left the decoder in `:start`; byte 1 was therefore reselected as a `0x80` UCS2 header and the result was `A` instead of GSM-mode `@A`. Expectation-first tests covered later `0x80`, `0x81`, and `0x82` octets under strict, discard, byte-substitution, callback, and every Stream split. A separate 256-byte record proves the same transition through the 255-byte admission boundary, including first-error ordering and callback offsets 0, 1, and 255. The focused RED ran 11 tests with 2 failures; its test-source SHA-256 before production changed was `3b520c45bc1e0363251fa1a0225163e87326da36808a946c9e89f08136d8aaff`. GREEN commits `:start` to `:gsm` only after recovery consumes a physical invalid default-alphabet unit. Framed UCS2 and structural-EOF paths retain their existing states and error kinds. The focused run passes 11/11, the complete package passes 161/161, production compilation is warning-clean, and formatting passes. Final source/test SHA-256 values are `d8b8cd490e3c9b6f496a581366c311621f11446a14ce9bb60ce41681e55e5690` and `3528f59b85e10559622913c496d54a4c75fcada0283c7fe899fb0e73a1c05829`. 54. PACKAGED CORE LINK RED/GREEN: an expectation-first package contract required the independently published README to link Iconvex through its durable Hex package page and reject the checkout-only `../iconvex` target. RED ran 8 tests with the one intended failure. GREEN replaces only that link with `https://hex.pm/packages/iconvex`; the same focused file passes 8/8. 55. MORSE TOKEN RECOVERY RED/GREEN: an expectation-first public recovery test supplied the invalid six-dot signal `......`. Native discard rejected the complete token, while callback discard consumed only its first octet and leaked the remaining five dots as the valid digit `5`. The focused RED ran 8 tests with the one intended failure; its test-source SHA-256 was `707e8c466f768beefb14672867568a99f31e9e5219fb5b93bb8cf586a4f45f11`. GREEN gives Morse invalid-sequence recovery the exact reported token width. Callback discard/replacement now emits one event and consumes all six octets, while byte substitution emits six `<2e>` fragments. The focused suite passes 8/8 and the complete package passes 163/163 with warnings as errors. Production compilation is warning-clean and formatting passes; the final Morse source SHA-256 is `ed09cf72a92e0f4ffc0272ae59483db3ceecffde3c38cd66456c55c08a2322e0`.