## avifenesh/memra — v0.137.0…v0.138.0

_134 commits._

### Features
- **prefill: prove which term of MTP walker support answered, and retire the short-first door on ornith's own receipt (#486)** (c4b0cfb)
- **dsv4: the matrix expert program is a NEW numeric class, with the drift rows it owes (memra #461) (#467)** (dc4bd19)
- **ctx-resolve: bank the red-arm demonstration for the five new tests** (09de92e)
- **feat: yield GLM5 prime chunks through the shared walker (#389)** (25c1f14)
- **feat: end the streaming session in Hebrew text, from the archive's own vocabulary** (ae4fd4e)
- **feat: read speech token ids back as text inside the engine** (ca3ceac)
- **feat: run a whole native streaming session, audio in and Hebrew tokens out** (1c9c3dd)
- **feat: decode Hebrew natively through the RNNT prompt, predictor, joint and greedy loop** (5f85cd3)
- **feat: run the RNNT log-mel natively and compose it with the encoder** (1d76c37)
- **feat: execute the cache-aware FastConformer encoder on the [56,0] streaming arm** (851206a)
- **feat: read and bind the Hebrew streaming RNNT checkpoint without running its pickle** (4a27c13)
- **feat: decode Whisper windows natively with the pinned beam-1 transcription policy** (8a6794e)
- **feat: match the Whisper frontend to the clip program real audio actually uses** (95620be)
- **feat: gate the native Whisper decoder and pin what each capture measured** (939930c)
- **feat: execute native Whisper encoder with matched precision gates** (fecc4fe)
- **feat: implement native Whisper log-mel with pinned CPU oracle** (e4c314c)
- **feat: add Whisper speech plan and strict source metadata pack** (d704aab)

### Fixes
- **fix(admit-predict): the boot line must name the arm it is in, and a test must read it (#477)** (622a346)
- **perf: restrict GLM TP pool-split indexer to prime (#395)** (9821f2c)
- **perf(serving): flip MEMRA_SERVE_DEVPENALTY default ON for Qwen3.5 on sm_120 (#429)** (3119489)
- **fix(gate): pass the new sp argument to gemma_spec_session_new** (496dfbe)
- **fix(clippy): index the gemma sampled burst's draft slots by iterator** (ce09ff9)
- **perf: default the wide norm2 pack ON with an explicit-0 seam** (caab94c)
- **perf: gate a wide same-class DSV4 norm2 pack (door MEMRA_DSV4_NORM2_WIDE)** (9e3787b)
- **fix: teach the shared replay census the norm2 door (memra #428)** (49dbb76)
- **perf: default the exact norm2 activation packing on** (6081d71)
- **perf: default the split-K paired-fetch entries ON with an unset/0 seam** (56f7683)
- **fix: repoint the releasability census teeth at the new publish anchor** (3491ee0)
- **perf: gate exact DSV4 norm and activation pack fusions (#422)** (4e10731)
- **perf: paired split-K loads with instrument + default-OFF door scaffold (#410)** (5ecd8b8)
- **perf: make the speech matrix product cache-blocked and optionally threaded** (ff68aa5)
- **perf: gate native Qwen GQA FA2 prefill on sm_120a (#421)** (9194395)
- **fix: keep encoder reduction bookkeeping clippy clean** (e8cafe4)
- **perf: default owner-accepted DSV4 HC S16 dots on (#418)** (fefb019)
- **perf: graph carried Qwen prime with live state tables (#417)** (4ea5de2)
- **fix: satisfy frontend lint and record restart status** (5259c54)
- **data: archive causal PMIN fix and remove unserved door (#393)** (950657b)
- **fix: include HC24 split in retained forward variants** (daf03e2)
- **perf: add default-off HC24 split dot numeric class** (a5c695c)
- **perf: fuse exact KDA verify input projections, save 1.790 ms/round (#388)** (cec4f5a)
- **perf: reduce Qwen prefill attention staging with exact bytes (#403)** (1864cde)
- **perf: default DSV4 dense-fast and norm-fuse on (#404)** (781db07)
- **perf: tune DSV4 dense exact-tree kernels (#396)** (8e8bd75)
- **perf: fuse adjacent DSV4 KV norm and RoPE behind a gate (#397)** (def42a6)

### Backend
- **release: v0.138.0 (#496)** (53c24cd)
- **speech: the 12 Hz decode step is launch-bound, measured, so step 1's priority stands (#494)** (6b500f7)
- **speech: strike implementation cost as a selection criterion, and say what the families are ranked on instead (#493)** (d71656c)
- **speech: TTS is a family in scope with a clock and a gate, and the concurrency leg was 0.53x the incumbent (#492)** (71e47db)
- **speech: the audio surface exists and sheds honestly, and the c4 collapse can be re-run in CI (#489)** (65ead73)
- **dsv4: the three norm-fusion doors are ported, measured on the served program, and deleted (memra #404, #426, #430) (#490)** (9474b75)
- **dsv4: an eos-terminated session parks at the boundary that includes the EOS it swallowed (#488)** (c540083)
- **dsv4: the matrix expert program is the default, and three doors leave with the flip (memra #461, #458, #460) (#482)** (fa92f54)
- **Delete the short-first prefill door, keep the rule it produced (#481)** (2817827)
- **research(glm5): the indexer-split lane guard grepped a decode shape a prime-only door cannot print (#475)** (ba850ae)
- **dsv4: count the dense entry points' own launch shapes, and name the per-row work (memra #472) (#473)** (3430700)
- **dsv4: the dense tile width becomes the code, and the door is deleted (#471)** (6fed745)
- **dsv4: dense wide-prefill tile width door, +4.10% served prefill, bit-equal (memra #463) (#468)** (94ead7c)
- **dsv4: a door on a program that cannot serve is permanently unreachable, not inert (#457, #454) (#465)** (4830a1f)
- **serve: admit by memory instead of a worst-case slot proxy (door OFF, decide-by 2026-09-23) (#431)** (ad02b8c)
- **dsv4: refuse matrix + default-ON split-K at load, and gate door defaults against the served path (#458, #454) (#462)** (a615118)
- **Merge pull request #455 from avifenesh/lane/ctx-resolve-from-config-20260910** (daef533)
- **ctx-resolve: tell the three checkpoint cases apart, and refuse the two that are not a window** (5c53c8c)
- **Merge pull request #444 from avifenesh/lane/devpenalty-qwen35-20260909** (05d8674)
- **server: resolve the context window from the model's config, never from a literal 8192** (0418366)
- **dsv4f: delete the split vocab head door on its own first receipt (memra #433) (#447)** (21e9a96)
- **Merge pull request #453 from avifenesh/lane/speech-k1-20260910** (3aef487)
- **Merge pull request #435 from avifenesh/lane/dsv4f-ar-phase-20260910** (8763f8b)
- **gate(ar-phase): pin the instrument door off in the three gate bins main added since this lane branched** (bcdfb2b)
- **gate(ar-phase): the AR epoch is the signal block's counter, so a site advances by the pool size** (601660b)
- **gate(ar-phase): pin norm2-wide in the measured-default list** (4775179)
- **dsv4 ar-phase: write down the closure check's blind spot instead of asserting it away** (0058186)
- **dsv4 ar-phase: the two clocks must agree that time passed, and the gate asserts the allocation's arm** (19e0d9d)
- **dsv4: causal phase instrument for the residual all-reduce pool (MEMRA_DSV4_AR_PHASE, default OFF)** (fb6a8ab)
- **qwen35: calibrated FP4-activation prefill program (A4), its gates, and the instruments behind a NEGATIVE verdict (#424)** (4f15d99)
- **gate: the dsv4 packs own their chat dialect in the engine, pinned on the serve-path render (#446)** (dabab6e)
- **Merge pull request #448 from avifenesh/lane/g7-asr-decode-20260910** (4a371c1)
- **Merge pull request #436 from avifenesh/lane/gemma4-full-serving-20260909** (63a1289)
- **Merge pull request #443 from avifenesh/lane/dsv4f-compose-three-flips-20260910** (53e5018)
- **gate: compose all three flipped DSV4F doors on one binary** (56e7c25)
- **Merge pull request #434 from avifenesh/lane/dsv4f-compose-flips-20260910** (c72b2a4)
- **gate: pin the norm2-wide door off in the two-door composition instrument** (13f6538)
- **gate: compose the two flipped DSV4F doors on one binary** (16f3f5c)
- **gate: resolve the graph split-K entry census from the door, not a symbol name (#438)** (8f2ace5)
- **Merge pull request #416 from avifenesh/lane/asr-modality-20260909** (e7f7596)
- **review: scope the parity claims to what the receipts actually derive** (eeff342)
- **Merge pull request #437 from avifenesh/lane/dsv4f-norm2-wide-sanitizers-20260910** (8fa3fdc)
- **serve: default the gemma sampled-spec admission ON** (925a7ce)
- **tokenizer: allow a DIRECTORY model to take its tokenizer from a verified GGUF** (58c97a7)
- **Merge pull request #430 from avifenesh/lane/dsv4f-norm2-wide-20260909** (4eaf708)
- **gate: correct which gemma_sample_gate arm the red arm falsifies** (b793c49)
- **gate: qualify the gemma sampled-spec admission with gemma_sample_gate** (f685593)
- **gate: count norm2-wide references apart from raw-bit comparisons** (667b283)
- **Merge pull request #426 from avifenesh/lane/dsv4f-norm-fuse2-default-on-20260909** (f4421cc)
- **data: close the 36-clip sweep, bank the stop receipt, stand on the partial verdict** (dec961a)
- **Merge pull request #425 from avifenesh/lane/dsv4f-splitk-fast-default-on-20260909** (bd4a6f5)
- **data: stop the sweep clean at 36 of 71 clips and clear both real divergences** (ed5ffe9)
- **wip: gemma sampled spec arm (rejection-sampling verify) + FLAGS row** (efc2155)
- **wip: gemma sampled spec arm** (4335b1d)
- **gguf: load the Gemma-4 12B Unified checkpoint on the dense pack** (28caf1e)
- **data: gate the streaming session on a second speaker and on silence** (451d94a)
- **data: gate the streaming session over the whole 15.8-second clip** (56f0c45)
- **data: correct the whatsapp-001 verdict, the native path was right and the reference was padded** (0a847c7)
- **merge: compose the ASR lane with current main** (ddd136a)
- **merge: compose ASR lane with current main kernel registry** (5e6ae5d)
- **data: bank negative verify graph receipts without inherited runtime (#402)** (70b9508)
- **Merge pull request #406 from avifenesh/lane/dsv4f-hc-dot-split-20260909** (c592095)
- **data: bank negative shared-expert dual receipts (#383)** (f900d80)
- **data: bank F16 rows numeric no-go receipts (#294)** (9fb394f)
- **data: bank MLA verify component and remove unserved door (#394)** (1582020)
- **research(qwen): close prefill GEMM lane with measured negative verdict (#411)** (eff0002)

### Tests
- **test: verify speech refusal and finish census provenance** (3503a3a)
- **test: add HC24 real component and numeric-class model gates** (e330db7)
- **test(dsv4): inherit default graph split-K in profile CLI (#407)** (1828196)
- **test: gate DSV4 dense-fast and norm-fuse composition (#401)** (d19d078)

### Docs
- **docs(releasing): the v0.138.0 row counted the battery base, not the tag (#497)** (1c0900e)
- **docs(flags): repoint the DSA named class receipt at a bar that could fail (#474)** (f15aed1)
- **docs: correct the TP spec door row — the cause is the DSA named class, not the verify walk (#469)** (8bbb458)
- **docs: the TP spec composition door records why it cannot be flipped (#466)** (ec4f3b7)
- **docs(speech): record the K1 verdict and reorder the plan behind it** (1b7b096)
- **docs(FLAGS): give the AR phase instrument its own section, three cells like its table** (d79676c)
- **docs(speech): keep the two binding gate rules in section 5, not inside 5.1** (14ed0a6)
- **docs(speech): G7 is built, the served ASR decode contract** (3c7730a)
- **docs(engine): keep the B200 posture receipts on the function they document** (bb46142)
- **docs: the speech capability program, its gate set and its kill criteria (#440)** (15ef97b)
- **docs: close the norm2-wide sanitizer gap and withdraw the instrument-floor reading** (18a750e)
- **docs(flags): cover the gemma_sample_gate MEMRA_GSG_* family** (d2ee440)
- **docs: point the norm_fuse2 FLAGS row at the flip receipts** (035ab93)
- **docs: report the eight-clip Whisper parity and score it against both oracle backends** (b068997)
- **docs: record the measured ASR ladder and the first end-to-end parity numbers** (02b095b)
- **docs: close redundant GLM KDA verify rows proposal (#386)** (31886bd)
- **docs: retain qualified HC24 S16 door with drift decision pending** (d12e588)

### Chore
- **ci(local-ci): MEMRA_CI_RELEASE_CARD=0 skip door for one-model contributor boxes** (c804283)
- **chore: publish memra-tokenizer before memra-reference** (7658ba1)
- **chore: remove unserved KDA verify fusion and keep receipts (#413)** (1657a5a)

_Recap by [Repo Wrapped](https://repowrapped.com/gh/avifenesh/memra?utm_source=github-action)._