docs(TODO.impl): static speed branch positive; engram module landed

TODO.impl/10-engram-lexical-memory.md

@@ -19,21 +19,54 @@ a hashed n-gram lookup table is a direct fit for the idiom memory a
 - Corrected mechanism per 09: hash-addressed tables in the forward
   pass (DeepSeek config), not example retrieval.
 
+## Gate: OPENED (2026-09-12)
+
+02 (teacher mixture) and 04 (trained-init lite) both closed negative;
+by this file's own pre-registered condition, the lexical-memory lever
+is now the remaining architecture move for the student-side residual
+(news/wiki-skewed, interaction-level).
+
 ## If built
 
 - Table as a parameter module (addresses = n-gram hashes of the input
-  bytes; embeddings summed at the two chosen layers); trained with
-  the distill recipe; Sinkhorn-balanced table updates (06) are the
-  natural pairing, as in the report.
-- Export probe first: a table-augmented student exported to ONNX with
-  decode-health smoke before any quality run.
+  BYTES - the corpus is byte-level, so orders {2,3,4} are byte
+  n-grams); embeddings summed into the encoder stream at ONE layer
+  (DeepSeek uses two at 552B; at 300M one is the proportionate dose).
+- Table updates pair with the Sinkhorn-balanced rule (06), as in the
+  report.
+
+## Size budget (honest arithmetic, int8 storage)
+
+| table | size |
+|---|---|
+| 1M entries x 64-dim int8 | 64 MiB |
+| 2M entries x 32-dim int8 | 64 MiB |
+| 4M entries x 64-dim int8 | 256 MiB (over budget) |
+
+The browser tier ships 95 MB today (lite int4). A 64 MiB table on top
+is a ~168 MB tier - a NEW tier between lite and 2.1-int8 (264 MB),
+defensible. Ceiling: ONE table, <= 64 MiB, int8-in-zip (fp16 at
+inference, dequantized at load: the zip stays small, RAM grows 2x).
+Primary config: 2M x 32 int8 (more lexical coverage per byte).
+
+## Export feasibility probe (before any training)
+
+Gather ops over an embedding table must survive torch.onnx.export ->
+ORT -> the IMF zip pipeline. Probe: tiny table (1k x 8) attached to
+the tiny T5 fixture, exported, loaded in ORT, decode-health smoke.
 
-## Acceptance (only after the gate)
+## Acceptance
 
-- [ ] Size/latency budget for the browser tier documented
-- [ ] Table module + export probe
-- [ ] One run, full-set verdict with intervals
+- [x] Size/latency budget documented (above)
+- [x] Table module (src/gpu/engram.py) + 6 unit tests (PR #215)
+- [x] Export probe: gather graph survives ONNX+ORT within 1e-4; the
+      hash becomes a per-runtime 20-line function (addresses as an
+      IMF input); int64 remainder addressing (primes available)
+- [ ] One distill run, full-set verdict with intervals (GPU spend:
+      owner call on the run, the module and probe are free)
 
 ## Status
 
-- [ ] Gated behind 02/04 results; not built
+- [x] Gate opened (02 and 04 closed negative)
+- [x] Module + probe landed
+- [ ] Run (owner-gated)

TODO.impl/11-static-int8-framing.md

@@ -57,6 +57,16 @@ published byte parity is unaffected; framing is the parity variable
 of mixed-shape serving (why the speculative tier degraded and was
 pulled — TODO.impl/01).
 
+## Speed branch: POSITIVE (2026-09-12, full-set)
+
+The static artifact's quality gate PASSED: **4.6241** full-set (n=1200)
+vs the shipped dynamic's 4.5701 — +0.054pp, the artifact-drift band —
+while carrying +8% CPU decode speed (78 vs 72 tok/s). Re-exporting all
+quantized artifacts through the static path is a release decision:
+modal_export gains the calibration stage, index-v6, golden regen, and
+the gate re-runs with --out so the delta ships with a CI. Point
+estimate recorded in RESULTS.md.
+
 ## Status
 
 - [x] Pre-registered (this file)

TODO.impl/README.md

@@ -23,8 +23,8 @@ Priorities:
 | 07 | diacritization-depth-knob.md | depth-conditioned vocalization | P3 | product decision |
 | 08 | golden-closeout.md | golden test PR, paper-c wording, release README note | P0 | closed 2026-09-12 (PR #211 merged) |
 | 09 | doc-corrections.md | Engram mechanism fix, mHC single-pass note | P2 | landed in PR #95 |
-| 10 | engram-lexical-memory.md | hashed n-gram memory module | P3 | gated behind 04 |
-| 11 | static-int8-framing.md | static scales as the framing fix | P0 | closed: node-kernel finding; static stays open on speed |
+| 10 | engram-lexical-memory.md | hashed n-gram memory module | P3 | gate opened; module + export probe landed (PR #215); run owner-gated |
+| 11 | static-int8-framing.md | static scales as the framing fix | P0 | framing closed (node-kernel); speed branch POSITIVE (4.6241, +8%) |
 
 Standing rules that apply to every item: full-set measurement or no
 claim; quantized parity is quality-level, not byte-level; LLM teachers