perf(monorepo): cache pkg_strategy + pkg_tags, batch ls-remote, prune dead walks

[BryanFRD]

Concrete perf followups from the gix migration audit. Cumulative ~1.5-6s on mono-large.

Items

• Cache `pkg_strategy` + `pkg_tags` per package (`src/monorepo/run/mod.rs:182-186, 253-256, 314-317, 563-566`). `pkg.effective_versioning(&workspace, &tags_for_package(&all_tags, &prefix))` is called 3–4× per package. `tags_for_package` is a linear scan over `all_tags` allocating a `Vec<&str>` each time. For 200 pkg × 4 calls × ~1000 tags = 800k iterations. Compute once at the top of the per-package loop and reuse. ~50–150ms saved on mono-large.

• `is_package_touched` quadratic over changed files × shared paths (`src/monorepo/util.rs:77-108`). Per-package iteration over `changed_files` with `f.starts_with(prefix)`. For 200 pkg × 500 changed_files × ~3 shared_paths = 300k string compares. Pre-sort `changed_files` once, then `binary_search` the prefix or use `BTreeSet::range(prefix..)`. ~20–100ms saved when `recover_missed_releases=true`.

• Cache parsed semver in TagIndexEntry (`src/git/tags.rs:93-125`). `find_highest_semver_tag` parses `semver::Version` on every entry in every call. For 200 pkg × ~1000 tags this is 200k re-parses. Add `parsed_version: Optionsemver::Version` to `TagIndexEntry`, populated once at `TagIndex::build` time. ~30–80ms saved.

• Replace per-tag `git rev-list -n 1` in push_tags (`src/git/push.rs:13-23`, called in loop at line 186). On Windows each `CreateProcess` is 10-20ms. For 200 tags pushed: 2-4s. Resolve via gix: `repo.find_reference(&format!("refs/tags/{tag}")).peel_to_id_in_place()` (already have a `Repository` handle). ~1-4s saved on push-heavy releases.

• Skip `collect_dirty_files` when no hook defined for the package (`src/monorepo/run/mod.rs:490, 672`). `collect_dirty_files` shells out to `git status --porcelain` once per package per hook context (2 contexts: post-bump + pre-commit). Gate by `resolve_hook(...).is_some()` — line 489 already does it for post-bump, line 671 doesn't. ~100ms–4s saved depending on hook usage.

• Move OrphanedTagStrategy::Warn early-return in find_matching_commit (`src/git/tags.rs:225-265`). When strategy is Warn (default), the function does the 1000-commit walk then returns None at line 259. Early-return before the walk. ~5-30ms per orphan, irrelevant in healthy repos but free win.

• Resolve `push_url` once and pass down (`src/git/push.rs:97, 172, 261, 288, 404` + `fetch.rs:10`). gix re-opens remotes on each call. ~5-15ms saved, trivial refactor.

Methodology

After each item, re-run `bench/micro` and `bench/competitive` on mono-large. Target: cumulative 100-300ms shaved off scan + 1-4s off push step. Track regressions on the micro benches that exercise these paths (`git_collect_tags`, `git_find_tag`, `full_check_flow`).

Related #471 (root perf N+1 audit), #476 (rayon parallelization — should land after these caching wins so the parallel work itself is cheaper).