fix(convert): mtime check before derive_project_slug (#612) — v1.3.74

[Pratiyush]

Summary

`convert_all` no longer calls `adapter.derive_project_slug(path)` on every session before the mtime check. On Codex CLI that helper opens every `.jsonl` to read the `session_meta` `cwd` field — so a no-op sync of a 5k-session corpus paid 5k needless file opens. The fix reorders the per-session loop so the cheap mtime check runs first; slug derivation, project filter, and ignore filter only run for sessions that actually need conversion.

Closes #612 (`#arch-h9`).

What changed

• `llmwiki/convert.py` — moved `path.stat()` + state-mtime check above the `derive_project_slug` call inside the per-session loop. No behavior change for sessions that DO get converted; the work is just reordered.
• `tests/test_convert_no_op_perf.py` — new regression test exercising a counting adapter that fails CI if a future change re-introduces the early-derive pattern.

What's new

Surface	Change
File opens per no-op Codex sync (5k corpus)	5,000 → 0
Mtime check ordering	`derive_project_slug → filters → mtime`
Test coverage	new `test_no_op_sync_does_not_call_derive_project_slug`

Behavioural delta

	Before	After
Per-session work for `state[key] == mtime` (unchanged)	`stat` + `derive_project_slug` (file open on Codex) + project filter + ignore	`stat` only
Per-session work for sessions that DO need conversion	identical (slug + filters + write)	identical (slug + filters + write)
Behavior change visible from outside	none	none
5k-session no-op sync wall-clock	(dominated by 5k file opens)	(dominated by stat() — negligible)

How to test it

```bash
python3 -m pytest tests/test_convert_no_op_perf.py -v
python3 -m pytest tests/ -q -m "not slow"

Manual perf sniff on a real Codex corpus:

python3 -m llmwiki sync # first run — converts everything
python3 -m llmwiki sync # second run — should print all-unchanged in <1s
```

Pre-merge checklist

• One intent — single perf-fix + corresponding regression test
• All CI checks green — local non-slow suite passes; CI to confirm
• Linked issue — `Closes high: #arch-h9 — convert_all is 300-line monstrosity with 3 separate write paths. CodexCliAdap... #612` in body
• Conventional-commit title — `fix(convert): ...`
• Tests added or updated — new `tests/test_convert_no_op_perf.py` asserts the contract; full existing suite passes
• CHANGELOG.md updated — `[1.3.74]` entry under Fixed
• Breaking changes flagged — N/A; no public surface change
• No new runtime dependencies — N/A
• No real session data — N/A; test uses an in-memory counting adapter
• No machine-specific paths — N/A
• Docs updated — N/A; the change is internal to the sync loop
• Release notes drafted — see CHANGELOG.md `[1.3.74]` Fixed bullet
• UI verified — N/A, no UI surface
• A11y verified — N/A, no UI surface
• Commits GPG-signed — yes
• Reviewer has read every changed line — diff is +176 / -12 across 6 files; most of the new lines are the regression test scaffold

Bundle

• `llmwiki/convert.py` — per-session loop reorder; `stat()`/mtime first, slug+filters second
• `tests/test_convert_no_op_perf.py` — new test using `_CountingAdapter` that registers temporarily, runs `convert_all` twice, and asserts `derive_project_slug` was not called on the second (no-op) run
• `llmwiki/init.py`, `pyproject.toml` — version 1.3.73 → 1.3.74
• `README.md` — version badge bump
• `CHANGELOG.md` — `[1.3.74]` entry under Fixed

Out of scope / follow-ups

• The issue title also called `convert_all` a "300-line monstrosity with 3 separate write paths". The two real write paths (markdown-source copy at the start of the loop, and parsed-jsonl render at the bottom) could be extracted into helpers in a future PR — held off here to keep this PR scoped to the perf bug. The reorder above is the part that was actually breaking on real corpora.

Next

After merge: tag `v1.3.74`, then run the README-claim audit (#682) or move directly to a final consolidated PyPI release (will need PyPI trusted-publisher to be configured first per the v1.2.0 plan; flagging that as the only remaining blocker).