v2.4.0 — parity release

Parity release, driven by an independent review of v2.3.1 (GPT 6 Astra, eight findings, all fixed; its ten reproducers are now tests/test_review_regressions.py) and by the gap to funes.

Added

• The commands Claude ran are recallable. Tool calls in a turn are captured as one compact line each ($ <shell command>, Edit <path>, Grep <pattern>, WebFetch <url>, …) into a new exchanges.tool_text column, indexed by FTS and shown under "Tools run". Tool output is still never stored. On a real 20 MB transcript: 141 of 163 exchanges carry tool calls, 327 shell commands.
• Ranked search with match snippets. Results are BM25 relevance re-weighted by recency (30-day half-life; --half-life 0 for pure relevance), best match first, and every hit carries the passage that matched (Match: …«term»…) instead of the first 200 characters.
• Secrets redacted at index time. Well-known credential formats (AWS, OpenAI/Anthropic, GitHub, Hugging Face, Slack, Google, JWTs, bearer tokens, PEM private keys), KEY=value assignments with secret-like names, and "my password is …" phrasings are replaced with [REDACTED:<kind>] before storage. Pattern-based, documented as not exhaustive (PRIVACY.md).
• Claude reaches for recall on its own. /recall:recall is now a skill (skills/recall/SKILL.md) with a trigger-phrase description, so Claude invokes it when the user refers to earlier work; a section tells it to search directly instead of showing the menu.
• Continuation-append indexer. Assistant blocks that land after their exchange was stored (a turn split across reads, a Stop-time transcript lag, a blocked-Stop continuation) are appended to that exchange, FTS kept in sync. Replaces the v2.3 hold-back rule and the last_assistant_message check.
• SessionEnd drains the whole backlog in committed passes within a 7s budget.

Fixed

• Retrieval hid stored answers after 1,000 characters (display used the user-side cap). Display caps now match storage (4,000) and the aggregate budget is 20,000.
• Capture could wedge forever on a turn larger than the 2 MB read budget, and on any single record larger than the budget. Every read now makes progress: at least one record is always consumed, and a prompt cut off by the cap is stored pending its reply.
• A partially written JSONL record could be skipped permanently (offset advanced past it). Unterminated, unparseable final records are no longer consumed.
• A proactive [Recall] suggestion discarded a due cross-session highlight in the same prompt; the /recall observability notice did the same. Outputs are now combined.
• Global search lacked the idx tiebreaker; ranking now ties on timestamp then idx.
• PRIVACY.md claimed credentials were never stored; they now are redacted, and the policy says exactly what that means.
• Hook commands left the plugin path unquoted (broken under paths with spaces).

Changed

• DB schema version 5 (tool_text column; FTS index rebuilt once).
• commands/recall.md removed in favour of the skill.
• 459 tests.