Cache Hatch workspace TOML parses by mtime in configure()

[Copilot]

Hatch::configure re-read and re-parsed pyproject.toml + hatch.toml for every workspace folder on every call. In the v1.33 insiders regression report, this per-call I/O (run while holding the configuration write lock) was the largest single amplifier of the cross-platform pet.refresh p90 regression.

Changes

• Per-workspace TOML cache in pet-hatch, keyed by (workspace_path, pyproject_mtime, hatch_toml_mtime). Cached entries are reused when both file mtimes are unchanged; missing files are recorded as None mtime so deletions invalidate correctly.
• configure() is now stat-only on warm cache — O(workspace_count) stats instead of O(workspace_count × 2) reads + parses. Cold path is unchanged.
• Cache lives behind the existing workspace_virtual_dirs mutex, so no new locking surface is introduced and ordering vs. lookups is preserved.
• Tests covering cache hit on unchanged mtimes, invalidation on TOML edit, invalidation on file deletion/creation, and behavior across multiple workspace folders.

Notes for reviewers

• Mtime resolution can be coarse on some filesystems (FAT, some network mounts). If a TOML is rewritten within the same mtime tick, the cache will return stale data until the next change. Calling out explicitly since this matches the issue's proposed design but is worth confirming acceptable.
• Branch was rebased onto main after fix(#461): release configuration write lock around locator.configure() #466 merged; this PR no longer touches jsonrpc.rs.

[Copilot]

Pull request overview

Adds an mtime-keyed parse cache to the Hatch locator so repeated configure() calls skip re-reading and re-parsing pyproject.toml/hatch.toml when those files are unchanged, addressing a pet.refresh p90 regression.

Changes:

• Introduces CachedWorkspaceConfig / ParsedCache and a parsed_cache field on Hatch, keyed by workspace path with (pyproject_mtime, hatch_toml_mtime) validation.
• Rewrites configure() to stat the two TOML files per workspace and reuse parsed results on a warm cache; cold path falls back to the existing parse.
• Adds tests for cache reuse, mtime-based invalidation, file-appearance invalidation, and eviction of dropped workspaces.

Show a summary per file

File	Description
crates/pet-hatch/src/lib.rs	Adds per-workspace mtime-keyed parse cache, updates configure() to use it, updates test helpers, and adds 4 cache tests.

Copilot's findings

• Files reviewed: 1/1 changed files
• Comments generated: 0

[eleanorjboyd]

Cache strategy is right: stat-only warm path, None mtime treated as a value so absence/presence transitions invalidate, eviction by rebuild rather than tombstones.

Two things to address before merge:

Publish atomicity — workspace_virtual_dirs and parsed_cache are assigned under separate locks. Wrap both publishes in one critical section, or document why the brief out-of-sync window is safe (only workspace_virtual_dirs is read by try_from, so external observers can't see the half-update).
TOCTOU comment — the "self-healing under TOCTOU" line overstates the guarantee. On coarse-mtime filesystems, two writes in the same tick won't bump the mtime; the stored value will then equal what the next configure observes and the stale parse will persist. Suggest softening to "the next configure with an updated mtime will re-parse" and leave the existing struct-level doc as the authoritative caveat.
On scope: keeping this Hatch-specific is correct. Every other locator's configure() is O(1) lock writes with no I/O, so there's nothing to cache there. Detecting Hatch before caching would require reading the exact files we're trying not to read — counterproductive.

[eleanorjboyd]

@copilot address comments

[Copilot]

@copilot address comments

Addressed in 36c901c: documented why the separate cache publishes are externally safe and softened the TOCTOU comment to only guarantee re-parse when a later configure observes an updated mtime.

[github-actions]

Performance Report (Linux) ✅

Metric	PR (P50)	PR (P95)	Baseline (P50)	Delta	Change
Server Startup	1ms	1ms	1ms	0ms	0%
Full Refresh	56ms	286ms	66ms	-10ms	-10.0%

Results based on 10 iterations. P50 = median, P95 = 95th percentile.

---

Legend

• 🚀 Significant speedup (>100ms faster)
• ✅ Faster than baseline
• ➖ No significant change
• 🔺 Slower than baseline (>100ms)
• ⚠️ Significant slowdown (>500ms)

[github-actions]

Performance Report (macOS)

Metric	PR (P50)	PR (P95)	Baseline (P50)	Delta
Server Startup	99ms	550ms	56ms	43ms
Full Refresh	155ms	34345ms	120ms	35ms

Results based on 10 iterations. P50 = median, P95 = 95th percentile.

---

Legend

• 🚀 Significant speedup (>100ms faster)
• ✅ Faster than baseline
• ➖ No significant change
• 🔺 Slower than baseline (>100ms)
• ⚠️ Significant slowdown (>500ms)

[github-actions]

Test Coverage Report (Linux)

Metric	Value
Current Coverage	79.9%
Base Branch Coverage	79.6%
Delta	.3% ✅

---

Coverage increased! Great work!

[github-actions]

Test Coverage Report (Windows)

Metric	Value
Current Coverage	76.96%
Base Branch Coverage	76.72%
Delta	0.24% ✅

---

Coverage increased! Great work!

[github-actions]

Performance Report (Windows) ➖

Metric	PR (P50)	PR (P95)	Baseline (P50)	Delta	Change
Server Startup	9ms	12ms	8ms	1ms	12.5%
Full Refresh	143ms	643ms	119ms	24ms	20.2%

Results based on 10 iterations. P50 = median, P95 = 95th percentile.

---

Legend

• 🚀 Significant speedup (>100ms faster)
• ✅ Faster than baseline
• ➖ No significant change
• 🔺 Slower than baseline (>100ms)
• ⚠️ Significant slowdown (>500ms)