feat(cli,store,common): Phase 66 Phase 4e — lpm cache prune + known-projects registry

[tolgaergin]

Summary

Closes the last Phase-4 prerequisite before the linker default flip (Phase 4f) can ship externally. The v2 virtual store at ~/.lpm/store/v2/links/ grows monotonically across every project on the machine — link entries from deleted/moved projects accumulate forever without prune. Phase 4e gives users the cleanup story.

Stacks on #33 (Phase 4d — silent migration + default flip). Will auto-rebase to main after #30 → #31 → #32 → #33 land.

Surface

lpm cache prune                            # dry-run, list orphans (uses registry)
lpm cache prune --apply                    # actually remove
lpm cache prune --max-age 30d              # only entries older than this
lpm cache prune --project <path>           # manual repair: ignore registry, walk this project
lpm cache prune --legacy-v1                # also wipe ~/.lpm/store/v1/ (post-migration cleanup)

lpm doctor now surfaces the orphan count and suggests prune as remediation when non-zero.

What's in this PR

lpm_common::known_projects — machine-global registry at ~/.lpm/known-projects.json. Schema-versioned JSON with canonicalized paths (so symlink-cwd quirks don't accumulate aliased entries). Atomic rewrites via <path>.tmp.<pid> → rename. Best-effort load policy (missing/malformed/schema-mismatch → empty registry, never blocks installs).

Install pipeline integration — every successful lpm install registers the project. Errors are logged + dropped at the call site so a flaky write never blocks an install.

lpm_store::v2::Store::iter_object_dirs() — pairs with the existing iter_link_entries for symmetric coverage of links/<*>/ and objects/<sri>/ orphans (preplan §4.4).

Prune algorithm (preplan §4.3 + §4.4) — collect roots from registry (or --project), BFS through LinkMeta.deps[].target_graph_key to mark reachable, apply --max-age filter, mark unreferenced objects via surviving link entries' object_path. Apply mode deletes; default is dry-run.

Doctor integration — V2_STORE_ORPHANS catalog entry, pass when zero orphans, warn-with-remediation ("run: lpm cache prune --apply") otherwise.

End-to-end smoke

$ lpm install                              # registers project in ~/.lpm/known-projects.json
$ lpm cache prune                          # → 0 orphan link entries, 0 orphan objects
$ rm -rf <project>                         # delete project
$ lpm cache prune                          # → 5 orphan links + 5 orphan objects, 9.4 MB eligible
$ lpm cache prune --apply                  # → Pruned 5 link entries + 5 objects (9.4 MB)
$ ls ~/.lpm/store/v2/links/                # empty

Pre-merge gate

cargo clippy --workspace --all-targets -- -D warnings  ✓
cargo fmt --check                                       ✓
cargo nextest run --workspace --exclude lpm-integration-tests
  → 5742 tests pass, 7 skipped                          ✓
cargo test -p lpm-auth (parallel-deterministic)         ✓
./bench/audit-fixtures/run-all.sh
  → 17 PASS / 1 SKIP / 0 mixed (v2 default)            ✓
LPM_STORE_VERSION=v1 ./bench/audit-fixtures/run-all.sh
  → 17 PASS / 1 SKIP / 0 mixed (v1 downgrade)          ✓

Tests added

• lpm_common::known_projects::tests (10) — schema versioning, malformed-JSON degrade, schema-mismatch degrade, symlink canonicalization, dedup-on-register, last_seen bumps, atomic no-tmp-leak, drop_missing happy path + zero-when-all-alive, sort-on-write.
• commands::cache_prune::tests (4) — orphan detection happy path, multi-hop BFS through dep edges (parent → child → grandchild), --max-age preserves recent unreachable entries, --project mode skips the registry.

Out of scope (queued)

• Concurrent install ↔ prune locking via lpm_root.store_lock() (production users running prune from a separate shell during an active install is the lone risk window).
• --gc-registry mode for collapsing very-stale (>180d) registry entries.
• Automatic doctor threshold tuning (currently warns on any orphan count > 0; preplan mentions "N GB / X% thresholds TBD").

These don't gate Phase 4f's linker default flip — that's the next milestone, gated on bench data.

Test plan

• All new unit tests pass (10 + 4)
• Full nextest workspace run passes (5742 tests)
• Audit fixtures green under both flag values
• End-to-end prune smoke (install → register → delete → prune dry-run → prune apply → store empty)
• Doctor reports orphans and suggests prune
• CI matrix green on both audit-fixtures (v1) and audit-fixtures (v2) rows

🤖 Generated with Claude Code