feat(badge/status/installer): fix badge, add status inventory, agy trust

[s-annam]

Summary

Three related okfmem changes landed together on one umbrella branch as a single
commit:

• fix(badge) Save-state badge false 'unsaved': scratchpad/temp writes count as work #49 — scratch/temp file writes no longer count as "work" toward
  the save-state badge, so a session that only touches throwaway files isn't
  nagged to save. Tightened during review (see below) so a project-internal
  docs/scratchpad/ can never be mistaken for the harness scratchpad.
• feat(status) okfmem status: add per-project inventory to the engine (page/archive counts, cwd->project, MEMORY.md limit) #43 — okfmem status gains a per-project inventory section
  (page/archive counts, MEMORY.md line count with the 200-line auto-load
  warning, cwd→project marker, decay: line), surfaced from the engine so
  /okfmem reports identically on macOS and Windows; skill loses its POSIX-only
  shell loop.
• feat(installer) Installer: Pre-grant okfmem-store path permissions for Antigravity (agy) in install.sh / install.ps1 #42 — install/uninstall scripts pre-grant Antigravity/agy
  trust for the store path via ~/.gemini/trustedFolders.json (least-privilege
  TRUST_FOLDER on the store abspath), gated on [y/N] consent, skippable
  non-interactively (prints okfmem init --grant-agy), with matching uninstall
  cleanup that removes only okfmem's own entry.

Resolves #49
Resolves #43
Resolves #42

Also note: #44 (structural save-state parser) was already shipped by #45/#47 and
is not part of this diff — it can be closed separately.

Test plan

• python3 scripts/check-leaks.py exits 0 (64 files, no leaks)
• python3 -m pytest -q → 246 passed, 1 skipped
• bash -n install.sh uninstall.sh OK; .ps1 files pure ASCII
• Windows CI matrix (verify) green
• Smoke: okfmem status / okfmem status --all render the new section

Adversarial review

An independent reviewer attacked the accumulated diff before this PR was opened.

Round 1 — 1 blocking finding (fixed): is_scratch_path's blanket
scratchpad-path-segment fallback matched any path containing a directory
named scratchpad at any depth — so a committed docs/scratchpad/adr.md edit
would classify as non-work, clear the badge, and silently lose a genuinely
uncaptured session (the exact failure the badge exists to prevent). Fixed by
dropping the blanket fallback: is_scratch_path now returns True only for
paths under _ephemeral_roots() (OS temp roots + $CLAUDE_SCRATCHPAD_DIR). The
real harness scratchpad already lives under /private/tmp, so it's still caught;
a project-internal scratchpad/ is not. Regression tests added; the strictly
narrower rule can only ever err toward a benign unsaved nag, never a false
saved.

Nits (not blocking): corrupt hand-edited trustedFolders.json is replaced
rather than merged on grant (mitigated — a .okfmem.bak backup is written
first); /okfmem no longer reports the engine clone's own git status (arguable
intentional scope-narrowing to store-backup health).

Reviewer verdict after the fix: #43 status-inventory and #42 agy-grant consent
gating were both found solid — consent complete on every rung-2 write path,
probes read-only and dispatched before the store-shape check, revoke touches
only okfmem's own entry, no argparse-flag collisions.

Provenance

Stage	Model	Effort
Implementation — #49	Claude Opus 4.8	high (--quality)
Implementation — #43	Claude Opus 4.8	high (--quality)
Implementation — #42	Claude Opus 4.8	high (--quality)
Adversarial review	code-reviewer agent (model not resolved)	—
Blocking-fix	Claude Opus 4.8	high
Orchestration	Claude Opus 4.8 (1M context)	—
Verification	CI verify	pending

[s-annam]

Windows testing

Pulled the branch and ran the full gate on Windows 11 / PowerShell — reproduced the red build (windows-latest) check locally, fixed it, verified green.

Root cause of the CI failure

tests/test_agy_grant.py hardcoded the POSIX store key /store/okfmem-store and compared against it raw. But grant_agy_store_access / revoke_agy_store_access / agy_grant_state all normalize the store path with os.path.abspath(os.path.expanduser(...)), which on Windows prepends the drive — D:\store\okfmem-store on the CI runner, C:\store\okfmem-store locally. Production code was correct; the tests were not portable (same class of footgun as the encode_root round-trip guard).

Fix

Added KEY = os.path.abspath(os.path.expanduser(STORE)) and compare/seed against KEY instead of the raw POSIX literal. Two tests (test_revoke_removes_only_our_entry, test_revoke_dry_run_writes_nothing) had been passing by accident — the raw-key lookup early-returned as a no-op, never exercising the revoke/dry-run path — so those were switched to KEY too and now test the real behavior. No production code touched.

Verified on Windows

• python -m pytest -q → 244 passed, 3 skipped (was 5 failed) — test_agy_grant.py 12/12 green
• python scripts/check-leaks.py → clean, 64 tracked files
• Smoke — okfmem status: add per-project inventory to the engine (page/archive counts, cwd->project, MEMORY.md limit) #43 status inventory renders correctly on Windows:
  • okfmem status — per-project pages:/MEMORY.md:/archived:/STATE: columns, * cwd→project marker on okfmem, ! over 200-line auto-load limit firing only where MEMORY.md > 200, decay: epoch … line, + N more truncation hint
  • okfmem status --all — all 14 projects listed, no truncation, aligned columns

CI after amend

build (windows-latest) pass, build (ubuntu-latest) pass, verify pass. Amended into the single commit (2b2504b → 38c05d4), force-push-with-lease, one-commit-per-PR preserved.