0.4.11 — Topic-aware retrieval ranking and per-occurrence tool binding

Release 0.4.11 — Topic-aware retrieval ranking and per-occurrence tool binding

Coordinated packages: ds4-context-core, ds4-context-reference-adapter, and ds4-context-engine 0.4.11.

Changes

• Stops treating generic workflow vocabulary as a historical query. devi, procedi, allora, aggiungi, continua, considerare, ovvero, quelli, fix, please and similar words no longer become keywords, symbols or query terms. Evidence that is explicitly present in the request — backticked spans, qualified identifiers, file paths, flags, error codes and quoted phrases — stays searchable even when its text is otherwise a stopword, so a legitimate topic named Continue or devi is still found. A bare acknowledgement now produces no-query instead of retrieving earlier approvals.
• Adds a bounded distinct-topic bonus of 40 per additional matched query term, capped at 80. FTS order, source role, recency and token length combined can no longer outweigh a candidate that matches one more topic term, while literal identifier and literal phrase tiers stay strictly ahead of numeric score.
• Counts matches with word boundaries and Latin accent folding consistent with the canonical unicode61 remove_diacritics 2 index. A keyword prefix inside an unrelated larger word no longer earns an extra match, and accented evidence matches its unaccented query term. Quoted evidence text is unchanged.
• Binds each tool result to the latest preceding occurrence of its tool-call ID. When a completion reuses an ID, independent turns are no longer merged into one oversized atomic group, an incomplete call can no longer be validated by a later reuse, and orphan results never bind to a future call. containsToolExchange is derived from completed call indices rather than rescanning every relation per group.
• Adds synthetic unit and integration regressions for acknowledgement queries, Italian topic extraction, stopword-named identifiers, workflow-only versus two-topic ranking, duplicate rows with a longer lower-ranked candidate, prefix and accent match accounting, reused IDs across independent turns, incomplete and orphan exchanges, a real exchange spanning a user boundary, a 200-exchange reused-ID sequence, a long native context with a 64k tail and an oversized immediate predecessor with input headroom. Each fails on the prior implementation and passes with the patch.

Implementation commit: c0298ab. Decision: ADR 071.

Scope and limitations

Both defects were reproduced locally from the code and its tests. Neither is confirmed to have occurred in a reported live session, and no live session, database or session JSONL was inspected for this patch.

No retrieval budget, result limit, token ceiling, recent-tail policy, recurrence ceiling, rescueImmediatePredecessor behavior, deduplication, branch isolation, evidence rendering, privacy handling or provider-facing default changed. context.recentTailTokens keeps its model-dependent automatic ceiling; the 64k ceiling applies only from a 1,050,000-token context window, not at 262,144. A genuine older oversized turn can still be excluded even when the whole native input alone would fit the global budget.

The ranking bonus is lexical, not semantic: it changes which historical evidence reaches the provider for a given request and cannot guarantee relevance. No live provider calls, real compactions or BPE/auto-tuning verification were performed, and the separate real-context objective remains open. Full real-context BPE/auto-tuning verification remains a separate objective.

Compaction validation, exact-value downgrading, compaction transport retries, graph atomicity, JSONL authority, persistence schemas and opt-in BPE/auto-tuning defaults are unchanged. Core and the reference adapter receive only the coordinated version bump.

Validation

On the final 0.4.11 source, version and dependency state, using Node 26.5.1 / npm 11.17.0:

• npm ci completed successfully.
• The approved jev_verify all gate passed npm run typecheck, npm test and git diff --check: workspace builds, strict TypeScript and the full Vitest suite, 103 files / 717 tests, including compatibility, privacy and integration tests. During development the new regressions were confirmed to fail on the unpatched implementation before passing with it. These are the build/typecheck/test scopes of npm run check, executed through the acquired verification policy rather than claiming a separate execution of that command.
• npm run quality:compare: candidate quality 0.9875 versus baseline 0.808156 on the packaged synthetic corpus.
• npm run schema:context-persistence: passed: true, 1,266 bytes / 317 estimated tokens, within the 320-token relative gate.
• npm run pack:check: verified core 251 files, reference adapter 7 files and engine 120 files in a clean consumer, including packaged exports, reference conformance and isolated offline Pi RPC.
• npm audit --omit=dev --json: no production findings. npm ci reports development-inclusive audit findings; the lockfile diff changes only coordinated versions and exact internal dependencies, not third-party package versions. These warnings are not hidden or fixed by this patch.

No live-provider, live-session or Windows-host confirmation is claimed.

Publication and registry verification

• Published manually to npm in dependency order: ds4-context-core@0.4.11 → ds4-context-reference-adapter@0.4.11 → ds4-context-engine@0.4.11, all public with latest.
• Annotated tag v0.4.11 and the GitHub Release were created after successful registry verification.
• CI remains validation-only; no publishing credentials, provider calls or session data were added.