perf(windows): Node hot paths for the check-inbox hook and CLI entries (implements #449)

[koujimatsumoto76-spec]

What

Implements the proposal from #449. On Windows, the measured hot hook / CLI entries today
walk bash -> many short-lived processes (sed / tr / sqlite3 / identity walk),
which costs 40-80 s per invocation on a stock Git Bash install (measured
breakdown in #449). This PR adds first-class Node implementations for the hot
paths and dispatches to them where it pays:

1. Identity resolver module (scripts/lib/identity-resolver.js) — pure
   lookup: project + type -> identity candidates. Reused by every entry below;
   replaces the per-call identities.sh spawn (~70 s of a 77 s SessionStart).
2. check-inbox.js — the Stop-hook path (cooldown, actas-lock respect,
   unread select, read_at on displayed rows, block/status JSON) via
   node:sqlite.
3. Fast CLI (agmsg-fast.js) — team / send / history / inbox /
   whoami, byte-parity with the bash output contracts (raw-compared; see
   Behavior notes). Only commands with an existing upstream .sh entry are
   included — no upstream-unused code.
   (Native SessionStart is deliberately NOT in this PR: open PR feat(windows): native Codex hook transport (queue + out-of-sandbox dispatcher) #305 already
   reworks the Windows hook transport across the same files
   (hooks-json/delivery/_delivery). Our measurements strongly support that
   direction — the Git Bash entry alone costs 6-9 s, and our local native
   SessionStart cut 77 s to ~1-2 s — so we would rather bring those numbers and
   review energy to feat(windows): native Codex hook transport (queue + out-of-sandbox dispatcher) #305 than collide with it. If feat(windows): native Codex hook transport (queue + out-of-sandbox dispatcher) #305 stalls, a follow-up PR
   can carry it.)

Dispatch contract (uniform): Node path only on Windows; each entry
(check-inbox.js and the fast CLI) probes node:sqlite and falls back to
the existing bash path only on a failed probe (exit 78), i.e. before any
side effect. The gating is
deliberately narrow: the bash paths remain the contract of record on every
platform, so the Node paths can be extended to other OSes later without this
PR taking that risk now. node:sqlite is still marked experimental upstream
in Node; because every native entry is probe-gated, an API change degrades to
the bash path instead of breaking anything. After
the Node implementation has started, a non-zero exit is loud and does NOT
fall back (a mid-run bash retry could double side effects such as a send
INSERT or read_at updates). Non-Windows platforms are untouched.

Measured (Windows 11, Git Bash/MSYS2, Node v25)

path	before	after
Stop hook check-inbox	50-80 s per turn	1.4-2.1 s
team / history / inbox / whoami	8-90 s (timeouts common)	0.3-1.6 s
send	8.1 s	0.47 s

Behavior notes

• Parity is enforced by 19 golden contract fixtures plus raw byte comparison
  of representative outputs (LF-normalized). read_at is set only on the
  rows actually displayed — this matches current main (check-inbox.sh /
  inbox.sh already restrict to displayed IDs), so no behavior divergence
  against the tree this PR targets.
• The package keeps its engines contract (node >= 14): the native paths are
  probe-gated, so an older Node simply keeps the bash path.
• node:sqlite needs Node >= 22.5; the capability probe (not a version
  compare) decides, and older/absent Node keeps the bash path.
• In restricted environments where MSYS2 cannot create shared memory
  (CreateFileMapping error 5), the native path is an availability fix, not
  just performance.

Not in this PR

• No new IPC, no Node port of watcher/bridge internals, no change to the
  launcher's role-session/thread authority or the request-file contract,
  and — matching current behavior — the codex SessionStart still skips the
  post-request cleanup/GC half (upstream's codex plug already exits 0 after
  the request hand-off); other types' cleanup semantics are untouched.
• No fix for Windows/Git Bash: agmsg_agent_pid never resolves (MSYS ppid walk cannot cross into native processes) — cc-instance.* is never written, actas locks always GC as stale, despawn reports no-live-lock #371 (instance/liveness semantics) — this only removes the ppid
  walk from the hot paths. delivery.sh status reports 'stale pidfile' for a live bridge on Windows (MSYS kill -0 cannot see native node PIDs) #415 is untouched here (our local install also has
  a native read-only mode/pairs fast path, but upstream has no wired entry
  for them today, so they are left out of this PR rather than shipped as
  unused code).
• The config.sh 3-level key defect found during this work will be filed as
  a separate issue (not yet filed; no existing upstream issue matches).

Structure (patchless boundary, against main 3f87d60)

This PR does NOT port our local patch artifacts (full-file replacements /
anchor edits were derived against v1.1.6 and are not portable — e.g. the
request file grew from 5 local fields to the current 3-field contract, with
identity authority moved into the launcher). Instead it implements the same
behavior first-class in the current tree:

• scripts/lib/identity-resolver.js + scripts/check-inbox.js +
  scripts/agmsg-fast.js as new files (no runtime deps, node:sqlite only);
• OS/type-aware dispatch added where the entry points already branch:
  hook generation (commandWindows) and the .sh entry heads;
• launcher internals, role-session/thread authority, watcher, and bridge are
  NOT modified.

Minimal seam (5 change areas — 4 runtime files + tests — per on-tree verification against 3f87d60):

file	change
scripts/lib/identity-resolver.js (new)	shared identity lookup
scripts/check-inbox.js (new) + scripts/check-inbox.sh head	Stop-hook Node path + Windows-only dispatch
scripts/agmsg-fast.js (new) + the 5 CLI .sh entry heads	team/send/history/inbox/whoami Node paths + dispatch
Node units + bats (Linux/macOS legs)	contract fixtures, dispatch exit-78 semantics

No changes to hooks-json.sh / delivery.sh / _delivery.sh — zero overlap with
open PR #305 (Windows hook transport), by design.

Fallback semantics: diagnostics to stderr and read-only probing are allowed
before falling back (exit 78); once the Node implementation has started, a
non-zero exit is loud and never falls back — a bash retry could double side
effects (a second send INSERT, repeated read_at updates).

Validation gate (before submission): the implementation is built and verified
on an isolated clone of current main. Nothing is transplanted from our
v1.1.6-based local install. Per-leg time boxes (a run is terminated at 5 min
of silence or its cap, and switched to an alternative gate instead of
retried): Node units <= 3 min; real commandWindows PowerShell smokes <= 60 s
each; static checks / bash -n <= 1 min; full Bats on Linux/macOS (CI or an
equivalent environment) <= 10 min. Windows Bats suites are NOT part of the
gate — measured on this machine they cost 4-7 min for 3 cases (delivery
subset aborted at 15 min), which is the very spawn tax this PR addresses;
the Windows delivery behavior is covered by the PowerShell smokes instead,
and that substitution is recorded openly.
Compatibility gates: shared-path shell code is bash-3.2-safe (macOS ships
3.2); a fixture pins the POSIX command field byte-unchanged when
commandWindows is added; full Bats pass on the macOS leg.

🤖 Generated with Claude Code