feat(core): topical dedup context — build_dedup_context + WASM/PyO3 bindings (PRD-04 DEP-3, Part A)

[p-diogo]

What

PRD-04 DEP-3 (F2), Part A: core only. This adds the new module rust/totalreclaw-core/src/dedup_context.rs. build_dedup_context(topical, pinned, recent, cap) -> String renders the "Existing memories" block the extraction prompt receives:

• the header is byte-identical to today's inline block in Hermes and the OpenClaw plugin;
• one [ID: <fact id>] <text> line per memory, pinned first, then topical, then recent;
• duplicate fact ids keep the first occurrence;
• at most cap lines (default 30); "" when nothing survives.

Bindings: WASM buildDedupContext, PyO3 build_dedup_context (JSON in, String out).

Parity

tests/parity/fixtures/dedup-context-v1.json holds 14 cases. generate-dedup-context-v1.py writes it from an independent Python reference implementation. Three suites check it:

• Rust: cargo test dedup_context (15 tests)
• WASM: tests/parity/dedup-context-parity.test.ts (new step in CI job cross-language-parity)
• PyO3: python/tests/test_dedup_context_parity.py (CI job python-tests)

Spec deviation

Enclave spec §4.6 sketches Vec<Claim>. The v0 Claim has no id, and each line must carry the on-chain fact id that the LLM echoes back as existingFactId. The element type is therefore DedupContextItem { id, text }.

Not in this PR

• No version bump. The coordinator opens the core 2.7.0 line.
• The Hermes adapter (agent/lifecycle.py) is Part B. It is dispatched after totalreclaw-core 2.7.0 is published.

🤖 Generated with Claude Code

[p-diogo]

Whole-PR review (Sonnet): verdict MERGE-WITH-FOLLOW-UPS (nothing blocking)

I read the code and the CI results. Rust Core, Python, Cross-Language Parity and Phrase Safety all passed; the TS, MCP, plugin and web jobs were legitimately path-filtered.

• Interface fidelity. build_dedup_context(topical, pinned, recent, cap) has the default cap of 30. It orders pinned → topical → recent, dedupes by id with the first occurrence winning, and stops at cap. Fixture cases pin the edge cases: cap=0, all sections empty, a duplicate id across sections, and pinned count ≥ cap.
• Header. It is byte-identical to python/src/totalreclaw/agent/extraction.py:1028/:1185 and mcp/src/extraction/extractor.ts:761/:1440, and the per-line format [ID: id] text matches. Part B keeps the leading "\n\n" (D2). Two behaviour changes, intended and belonging to Part B: the cap drops from 50 to 30, and ids and texts are collapsed onto one line.
• CI ordering. The new step runs in cross-language-parity after the wasm-pack build (ci.yml:295).
• Release. No version or release files are touched, and the CHANGELOG entry sits under [Unreleased]. The 2.7.0 release is the coordinator's (Task A4).

Follow-ups (tracked in the internal repo):

1. The module and function docs should state the caller contract: per-section limits (top-20 / last-10) are the caller's job, and once the pinned count reaches cap, pinned facts crowd out topical and recent ones.
2. A negative cap from JS coerces to a huge usize (effectively uncapped), while PyO3 raises OverflowError. Add a one-line guard or a doc note.
3. Add a PyO3 ValueError test for non-array input.
4. Python str.strip() also strips U+001C–U+001F and Rust trim does not. This is reachable only through the fixture generator's reference; production runs through Rust. Fix the generator the next time it's touched, or add a fixture case with \x1c at a line edge.