fix(windows/capture): production hardening from a real 3-project deployment (closes #1, closes #2) + opt-in --global single-daemon mode

[rui08984-dot]

Context

This branch came from running SessionRelay 0.5.1 as the memory layer for three real projects on one heavy Windows machine (i7-14700HX, 3 project roots sharing one ZCode source db, ~25k captured messages). Two red-team rounds + a full source audit surfaced the root causes behind #1 and #2. Everything below is verified against your own test suite: 222 tests → 221 pass + 1 skipped, 0 failures (also defused a time-bomb in test/forget/functional.spec.ts whose absolute fixture date broke its own assertion past a 30-day boundary — the suite is fully green for the first time).

Fixes

#1 — Windows boot chain (src/cli/service.ts)

• windowsRunScript: quote nodeAbs (C:\Program Files\nodejs contains spaces → 'C:\Program' is not recognized), and emit chcp 65001 >nul first — cmd.exe parses the file in the OEM codepage (GBK on Chinese systems) while we write UTF-8, so CJK project paths corrupted cd /d and the log redirect.
• vbs is now written UTF-16LE with BOM: wscript only reads ANSI/UTF-16; UTF-8 vbs with CJK paths misreads as GBK (one layer deeper than the cmd bug).

#2 — daemon resource/lifecycle (src/capture/watch.ts, src/cli/watch.ts, src/adapters/zcode/index.ts)

• Root cause of handle leak + 1GB memory sawtooth: detectCompaction opened a new better-sqlite3 connection per session per cycle and never closed it (the finally was empty — a stale "cache connection" comment copied from getConn). Now reuses getConn. healthCheck closes its probe too. Production evidence: handle slope ~378/h → flat.
• Single instance: the file lock (check-then-write) races under concurrent spawns; production accumulated 18 instances. Added a per-project named-pipe lock (\\.\pipe\srelay-watch-<sha1(root)[:16]>) — kernel-owned, dies with the process, EADDRINUSE = already running. win32-only; other platforms unchanged.

Capture correctness / quality

• extract.ts: decision extraction matched trigger words anywhere in a sentence and captured kanban table rows, markdown bold fragments and embedded-JSON residue. Added a plausibility guard (min length, trailing colon/pipe/bold rejection, table-row rejection, JSON residue rejection).
• bin/srelay.ts unresolved: --json existed but was dropped in the action — always emitted human text. Now passed through.
• cli/meta.ts decisions --limit: took the oldest N (ascending sort + slice(0, n)) while labelled "recent" — now takes the newest N; unresolved gets a 14-day fade.
• capture/sync.ts runSync: added a per-session content-signature watermark (changeProbe: zcode = two aggregate MAX(rowid) queries per cycle; file sources = mtime+size) plus a "session still in db" guard (so rebuild/forget flows still re-ingest) and a 10-minute full-sweep backstop. Idle daemon CPU went from ~7.5% of a core to ~0.1%.
• store/db.ts insertMessage: FTS index text capped at 100KB (jieba sync tokenize of multi-hundred-KB tool dumps stalled the event loop; full content is still stored).

Other warts found in the same audit

• capture/sync.ts: ensureRegistered called twice (second call returns empty — custom adapter errors were never reported).
• capture/archive.ts: --days 0 / --size 0mb treated as unset (falsy) → cutoff degraded to 9999-12-31, "archive nothing" archived everything (with --hard).
• shared/config.ts: corrupt config.json produced a raw stack in every CLI command → readable error with repair hint.
• cli/forget.ts: corrupt preview snapshot silently skipped the optimistic-lock diff before --yes → corrupt snapshot now refuses execution (fail-safe).
• mcp annotate_session / save --tag: rewriting meta_text dropped decision texts → viaMeta search surface silently narrowed; now re-included.
• adapters/claude-code/watcher.ts: runtime error events from fs.watch had no listener — a deleted/overflowed watched dir threw and killed the whole daemon; now degrades to the existing 1s polling fallback.

Opt-in feature: srelay watch --global

One daemon process adopts all projects from the existing registry (registryCandidates()), spawns per-project workers (each still holds its own per-project pipe/file locks), re-scans every 10 min to adopt new / prune removed projects, and dual-writes worker logs to each project's own watch.log (per-project --status tails stay fresh). Guards against a second global instance via a fixed-name pipe. Deployed on the 3-project machine: 3 daemons → 1, idle CPU ≈ 0, cross-project isolation verified.

Notes

• Windows-only paths degrade gracefully (pipe lock and chcp are win32-guarded; polling fallback unchanged for Linux).
• better-sqlite3 connections all carry busy_timeout already, so daemon+CLI+MCP concurrent writes are safe.
• Happy to split this into smaller PRs if you prefer — the commits are logically grouped (see the fork's history), and the --global part is independent of the fixes.

Closes #1, closes #2.

[rui08984-dot]

Follow-up: 3 more commits since opening (live field report from a 4th deployment)

A 4th project (E:/mc整合包, MC modpack) was onboarded on 09-23, and its AI operator did direct SQLite forensics (node:sqlite, bypassing the CLI) — the report caught an architectural defect the earlier rounds missed:

P0 — decisions froze for 20+ hours on long-lived sessions. Extraction only ran at the confirm transition (idle 10min → pending → 6h cooldown → confirm); a continuously-worked session never completes that loop, so sessions.decisions froze at first extraction. Fix (the reporter's suggested 1a): refreshExtraction() now runs at every pending transition — decisions lag ≤ 10 minutes. confirmSession reuses it.

P1 — 88% of expanded decisions were workflow-subagent fragments. The zcode adapter now tags title-matched subagent sessions with originHint='workflow' → origin column → listDecisions excludes them by default (data preserved, queries clean). Live on the 4th project: 36 sessions auto-tagged by the new code, visible decisions 82 → 20, residue 0.

P2 — decision text hard-cut at 90 chars produced mid-sentence / trailing-bracket fragments: cut at the nearest pause char + trim trailing closers + reject leading-closer fragments.

Plus: applyExtraction re-includes existing user_tags in meta_text (third site of the meta-rewrite family — annotate/save were fixed earlier, the confirm/refresh path was missed); listUnresolved dedupes the same question captured across sessions (briefs showed it twice).

Commits: 3e7679e (P0-P2), f17f52d (user_tags), 3b69246 (dedupe). Suite still 222 → 221 pass + 1 skipped, 0 failures. All changes verified on the live 4-project deployment (heartbeats, capture latency, brief contents).