Deduplicate project indexes by git/path identity

[Andy11-cpu]

Summary

• Add project_resolve module to match repos by git canonical root or canonical path
• Reuse existing cached project names when the same repo is opened from a different path
• MCP session detection and pipeline creation both use the resolver

Motivation

Split from #625 — prevents duplicate .db files when the same repository is indexed from multiple paths.

Test plan

• test_project_resolve.c (canonicalize, identity key, find existing, pipeline reuse)
• make -f Makefile.cbm test (CI)

Signed-off-by: Andy11-cpu canada11@duck.com

[DeusData]

Thanks for splitting out project identity deduping. Triage: high-priority path/project identity fix, but the PR is currently conflicting.

Please rebase on current main before deeper review. The key review point will be avoiding project-name drift while still deduplicating symlink/path/git-root identities, especially around relative paths and existing cache entries.

[DeusData]

Thank you — the problem is real (duplicate project indexes for the same repo) and splitting this out of #625 was the right move. After a careful review we can't merge the current identity key, though, and want to be specific about why: canonical_root is derived from git's common-dir and is deliberately identical for a main repo and all its linked worktrees (that's what it was built for in #749 — branch-level identity). Used as project identity, indexing a worktree resolves to the main repo's project and reuses/overwrites its DB — silently breaking multi-worktree workflows (which we use heavily ourselves). Two more issues: identity_nested in the reverse direction means opening a monorepo root adopts a child project's name and its incomplete graph; and the scan opens every cached .db + runs a git subprocess per project on every pipeline creation and session-detect, which is costly and invasive (read-write opens of other projects' stores). CI is currently red on all five platforms and the branch conflicts with main. Concrete path to merge: (1) key identity per-worktree — canonicalized worktree root, or (canonical_root, worktree_path) pair — so worktrees stay distinct projects; (2) drop the reverse-nesting adoption; (3) cache the identity scan (or restrict it to index-time, not every session-detect). Also note #625 declares itself gated on this PR, so it inherits the same rework. If you'd rather not carry the redesign we completely understand — say the word and we'll distill the per-worktree variant with credit to you. Thanks for pushing on a real pain point.

[Andy11-cpu]

Rebased on current main and reworked per your review. Summary of what changed:

1. Per-worktree identity (not canonical_root)
cbm_project_identity_key now uses the canonicalized git worktree root (rev-parse --show-toplevel + realpath/_fullpath). Linked worktrees keep distinct identity keys even when canonical_root matches. Symlinks / alternate paths into the same worktree still dedupe.

2. Dropped reverse nesting
cbm_find_existing_project_name only matches when the query path equals or is a subdirectory of an indexed root. Opening a monorepo root no longer adopts a child project name/DB.

3. Cached identity scan
The cache-dir scan runs once (read-only via cbm_store_open_path_query) and is held in memory until cbm_project_identity_cache_invalidate(). Invalidated after a successful full or incremental cbm_pipeline_run. Session detect still calls the resolver but hits the cache after the first load.

4. CI / conflict fixes

• Resolved Makefile.cbm conflict (test_index_resilience.c + test_project_resolve.c both kept).
• Fixed test_project_resolve.c compile errors (helpers, no ASSERT in void callbacks).
• Added project_resolve_worktrees_distinct test.

Commit: 9376ca1 on fix/project-dedup. CI should be re-running now.

[DeusData]

Thank you for tackling this — index duplication is a real annoyance and the git-worktree-identity idea is a good instinct. After a careful review we're going to close this one, and I want to be transparent about why:

1. The safe cases are already handled. Normalize project paths for ADR lookups #780 (merged) canonicalizes the repo path inside project-name derivation (realpath), so path-spelling variants (trailing slash, ./, ../) and symlinked spellings already resolve to the same project + .db — no duplicate index.

2. The one genuinely-new behavior carries a data-loss risk. The remaining novel case here is 'indexing a subdirectory of an already-indexed repo reuses the whole-repo project.' Via identity_is_child, indexing a nested/submodule repo (e.g. /repo/packages/api under an indexed /repo) reuses the parent's project name → the new repo resolves onto the parent's .db → the incremental-mismatch path cbm_unlinks and rebuilds it, destroying the original project's index. It's enabled by default with no nested-.git guard, and it also blocks a user from deliberately indexing a subdirectory as its own scope.

3. There's also a build issue (project_resolve.c:40 uses an undeclared buffer and writes to the wrong variable, so dedup is a no-op on Linux/macOS; CI's test job was skipped behind the failing lint/security gates).

Given #780 already covers the safe dedup and the new behavior is the risky part, we don't want to ship the auto-merge-and-delete identity model. If you'd like to pursue a safe version — detect a likely-duplicate index and warn (or offer an explicit, opt-in merge) rather than silently merging + unlinking — that'd be a welcome, much lower-risk direction and we'd happily review it. Really appreciate the effort here.