Make mnemon reachable inside the sandbox via the opt-in EXTRA_* mechanism

[audreyt]

Summary

Makes mnemon (recall + remember) reachable inside the sandbox via the general opt-in OMP_SANDBOX_EXTRA_WRITE / EXTRA_READ mechanism — the same knob users already use for any other extra writable/readable path.

On the real-dir-at-~/.mnemon setup (the symlink-into-~/.local era is over), the carve-out was functionally redundant with EXTRA_*: both produce byte-identical (subpath "$HOME/.mnemon") rules, and the per-$HOME-ancestor stat() walk that justified the ancestor-literal lines only applied to the symlink setup.

Behavior change (intended, opt-in)

mnemon is no longer automatic. A user who wants recall/remember inside the sandbox sets:

export OMP_SANDBOX_EXTRA_WRITE="$HOME/.mnemon"
export OMP_SANDBOX_EXTRA_READ="$HOME/.mnemon"   # or inherit from WRITE

EXTRA_READ must cover every EXTRA_WRITE path: Seatbelt file-write* and file-read* are independent classes — a path in EXTRA_WRITE but not EXTRA_READ is write-allowed-yet-read-denied, which breaks SQLite WAL and read-modify-write patterns. The ~/.zshrc shipped with this repo sets EXTRA_READ to inherit EXTRA_WRITE, which is the safe default.

MNEMON_DATA_DIR env forward and the OMP_SANDBOX_NO_MNEMON opt-out are dropped — neither is needed once the store is a real dir at the default path (mnemon resolves ~/.mnemon itself given $HOME, which the existing home-node metadata rule already covers).

Self-test hardening

The previous "fold ~/.mnemon into EXTRA_* then gen_profile" was a no-op in-process (EXTRA_RULES computed once at script entry; re-exporting the env var does not recompute the shell var gen_profile interpolates), and test 15 only passed as a side effect of test 13's clobber-then-restore of $PROFILE via a fresh --print-profile subprocess — reorder the tests and it would silently break.

• _sbx_build_extra_rules() refactored into a function, called at top level and re-callable from the self-test.
• Test 15 made self-contained: generates its OWN profile via a fresh "$0" --print-profile with EXTRA_*=~/.mnemon, then asserts recall returned a result (greps JSON for "id"), not just exit 0 — mnemon recall can exit 0 with empty output on failure modes, which would otherwise be a false pass.
• Test 13's restore now restores $PROFILE to the pre-canary entry state (captured _entry_x*), no longer touching mnemon.

Verification (truly clean env — env -i, no EXTRA_*)

• bash -n and /bin/bash -n (Apple stock bash 3.2): clean.
• --self-test: 21 PASS. Lone FAIL is the pre-existing, unrelated --auto-approve flag accepted by omp (clean env) (omp version issue, not mnemon).
• Test 15 detects its defect: a control with test 15's profile forced to lack mnemon → test 15 prints FAIL (no false pass).
• Test 15 is order-independent: reordered before test 13 → still PASS.
• Production e2e (EXTRA_*=~/.mnemon): --print-profile shows ~/.mnemon in write+read; mnemon recall jensen returns 1 real result. No phantom ~/.local/share/mnemon recreated.

Files

• omp-sandboxed | 132 +- — refactor + self-test restructure + header comment refresh.
• README.md | +7 — reframe write/read allowlist tables to "opt-in via EXTRA_*"; add a mnemon example.

Draft vs review

Marked ready for review. maintainer_can_modify is enabled so you can adjust on your end if needed.