2026-07-31 ticket initialization

User decision: kontynuuj
Previous recommendation: replace further dictionary growth with a
language-independent topic-matching experiment.
Workflow state: TOOLS

Current known gap:
Kolejka zadań powinna ponawiać nieudane próby z opóźnieniem
src/queue/task-retry-backoff.ts
Result: 0/1 relation because lexical topics do not cross the language boundary.

Constraints:
offline CI remains provider-independent
three-topic hard-negative boundary remains in force
model-derived evidence must be explicit and auditable
external inputs remain tracked-only snapshots

2026-07-31 local embedding benchmark

MiniLM revision=86741b4e3f5cb7765a600d3a3d55a0f6a6cb443d
positive_min=0.673289 negative_max=0.732568 separation=-0.059279
pairwise_correct=5/6

E5 revision=f470c6a1a906014160ece1968c484b275f0396de
query_prefix=query: passage_prefix=passage:
positive_min=0.759374 negative_max=0.835202 separation=-0.075828
pairwise_correct=6/6 minimum_pairwise_margin=0.007190

Decision: no global cosine threshold is safe.

2026-07-31 tracked platform ranking

commit=3e96573d587cb664741849ceba205bf303b9f418
graph=ae92ead72d35e88e6de754d7af02d074201213c56166eb788a5e152d6a6f695d
module_aggregates=133 actionable_targetless_declarations=66

forward score>=0.75 margin>=0.01:
selected=6 new_candidates=2 manually_accepted=0

reciprocal top-1 with forward/reverse margin>=0.01:
selected=1 new_candidates=0

Decision: reject production embedding matcher; workflow TOOLS -> ANALYSIS.

2026-07-31 gold cohort

cross_language_cases=7
known_positive_relations=6 satisfied=0
forbidden_pairs=6 violations=0
gated exact-target/capability-topic precision=100% recall=100%
gold v1=PASS gold v2=PASS
Workflow transition: ANALYSIS -> VERIFY.

2026-07-31 full validation and close

$ npm run verify
tests=244 pass=243 fail=0 skip=1
Java skip reason: local JDK unavailable; required CI supplies JDK 17.
LLM boundary=PASS (9 entrypoints, 31 modules)
module graph=PASS (94 modules, 429 imports, no cycles)
environment contract=PASS (63/63)
workflows=PASS
generated analysis=PASS (18 files)

$ npm run evaluate:gold && npm run evaluate:gold:v1
Result: PASS, gated precision/recall 100%, stability PASS.
Cross-language: expected=0/6, forbidden violations=0/6.

$ npm run examples:check
Result: PASS, five SDK fingerprints agree.

$ npm audit --omit=dev
Result: PASS, 0 vulnerabilities.

$ make smoke protocol-smoke
Result: PASS, CLI, MCP and A2A.

$ make docker-smoke
Result: PASS.

Result: all acceptance criteria satisfied; workflow VERIFY -> DONE.

2026-07-31 repository layout correction

User review identified executable code under project/ticket-004.
Moved:
project/ticket-004/evaluate-embeddings.py
-> scripts/research/evaluate-embedding-pairs.py
project/ticket-004/rank-graph-embeddings.py
-> scripts/research/rank-intent-graph-embeddings.py

Benchmark inputs, captured results, decisions and raw logs remain under the
ticket. No experiment code remains there.