feat(boatstack): retromine — offline recurrence detection over transcripts

[bigboateng]

Seventh slice of the published → merged program (follows #241) — the first half of Phase D. The program's general law: a recurring operator instruction is steady-state error — evidence the controller is missing a typed observation, verb, setpoint, or guard. This PR builds the sensor for that law.

What changed

New pure package internal/retromine:

• Neutral event schema (JSONL: source, session_id, ts?, role: operator|agent|tool, text, index) any tool can emit — the miner is agent-agnostic by contract.
• Adapters for Claude Code session JSONL and prefix-annotated plain text (User: / Agent: / >), plus the neutral format, with per-file content sniffing. Lossy projection is one-directional by design: unknown entry shapes are skipped (host formats grow constantly), but an unparseable line is a typed error naming its line — mis-parsing can never silently become "no recurrence found".
• Deterministic recurrence detection: normalize (strip fenced payloads — pasted logs are payload, not instruction), token 3-shingles, Jaccard ≥ 0.6, greedy clustering in sorted order. Recurrence = ≥ 3 occurrences across ≥ 2 sessions (repetition inside one conversation is conversation, not steady-state error). Sub-4-token acknowledgements never enter the pool. Per-session event indexes are parser-assigned, so caller concatenation order cannot change the report.
• No user-facing verb yet — retro derive (classifier + proposals + CLI) is the next slice.

Boundary conformance

• Boundary: operator transcripts entering an analysis pipeline that ships publicly.
• Control law: retro-derivation-is-offline-and-deterministic — no network, no subprocesses, no filesystem, no clocks, no randomness; identical inputs in any order produce identical output.
• Required evidence: retromine_conformance_test.go — positive (planted 3×/2-session instruction clusters; both adapters project equivalent content to equivalent instruction streams; fenced-payload variance does not split clusters), negative (2 occurrences or 1 session never clusters), relation (input order cannot change the report), failure-state (malformed lines are typed errors), bypass (the package's import list is conformance-tested to grant no I/O capability at all — the boundary is structural, not behavioral), plus a golden report over synthetic fixtures.
• Failure behavior: parse errors are typed and positional; nothing is ever partially, silently mined.
• Release condition: test fixtures are synthetic only (everything under product-engineering-loop/ ships verbatim to the public repo); no real transcript content, paths, or session IDs anywhere.

Evidence

• go build ./... && go vet ./... && go test ./... — pass.
• Release note added: 2026-07-28-retromine-recurrence-detector.md.

Final slice next: retro derive — gap classification (observation/verb/setpoint/guard) + reviewable proposals, derive-only.