feat(core): add first-class OMP session indexing

[Dustella]

What and why

Closes #81.

OMP (Oh My Pi) stores Pi-compatible JSONL v3 sessions under ~/.omp/agent/sessions, independently from Pi's ~/.pi/agent/sessions. Reusing the Pi source identity would lose provenance and prevent both histories from being indexed together. OMP also writes a mutable title record before the canonical session header.

This change adds a dedicated omp provider and parameterizes the existing Pi projection internally. OMP therefore reuses the established tree, branch, compaction, tool, usage, visibility, cursor, and raw-evidence semantics without copying the parser. The OMP-only configuration consumes the title prelude as metadata while preserving the composite session identity. Pi behavior and storage identity remain unchanged.

The provider is registered in the built-in catalog, watcher roots, Core package exports, README, skill, and schema/API references.

Verification

• npm test — 468 pass / 0 fail
• npm run typecheck — 0 errors (root + app)
• npm run lint — 0 errors / 4 pre-existing warnings
• npm run test:electron:all — not required; no app/ source changed
• New tests actually run in CI (.github/workflows/) — the current workflows only run the CLI subset; the repository-wide suite was run locally
• No existing assertion was loosened; provider catalog/watch-root/schema assertions were extended for the new source
• Every capability described above was exercised end to end from the provider entry point
• Re-ran the checks above on the current head after syncing upstream/main

Additional evidence:

• The OMP fixture preserves real OMP wire-record shapes with project content, paths, and signatures redacted.
• discover() is exercised against OMP's project-directory layout.
• Direct canonical assembly equals the SQLite round-trip result.
• A read-only smoke test against a local ~/.omp/agent/sessions corpus discovered 85 sessions and verified OMP provenance, title metadata, tool calls, and raw evidence.

Deliberately out of scope

• Kiro CLI uses a different session format and needs a separate provider.
• OMP custom session-root discovery beyond the default is not inferred; callers can configure an absolute provider root.

---

Provider adapter

• Read the neighboring provider implementations and reused the registry/projection contracts
• Session identity combines the normalized cwd namespace and header id
• A test calls discover() against the supported directory layout
• Directory layout was verified against OMP's emitted session corpus
• Canonical transcript invariant holds: direct assembly equals SQLite round-trip
• Text-less assistant records remain canonical message rows
• OMP has an independent indexVersionMarker
• Superseded history uses visibility
• The inherited snapshot cursor detects same-millisecond rewrites
• Version handling remains the existing Pi-family behavior

[tommy0103]

Reviewed the current head (12b4697). The OMP provider itself looks good — the main feedback is about scope, plus one query-layer bug.

Suggest splitting into two PRs

This PR bundles two separable changes:

• Commits 1–3: the OMP provider (the actual #81 work) — clean, well-tested, ready to merge on its own.
• Commit 4 (fix(query)): provider-agnostic query enhancements (projectPath, excludeInvoking, contextLimit, CLI --compact/--limit/--context/--snippet-length). Useful, but does not depend on OMP.

All open issues below live in commit 4, so the OMP provider is currently blocked by unrelated work. Commit 4 touches almost disjoint files; the split should be a trivial cherry-pick, and the two PRs can land in either order.

Blocking (query commit): excludeInvoking drops unattributed memories

buildWhere generates mem.session_id != ?. remember() stores session_id as NULL when omitted, and NULL != x evaluates to NULL in SQL, so memories({ excludeInvoking: true }) silently filters out every memory without a source session — verified against node:sqlite. Suggest (mem.session_id IS NULL OR mem.session_id != ?) scoped to memories (the other tables' session_id is non-null in practice). The new tests don't catch this because every fixture memory carries a session_id.

Conflict with main

main has moved to #120; skill-doc/SKILL.md now conflicts on the --search example line (main changed the nonce example; this PR adds the compact flags to the same line). Needs a rebase.

Docs / language consistency

• Commit 4's message is in Chinese; the repo convention is English conventional commits.
• New Chinese prose in SKILL.md, the api-reference.md tables, and packages/cli/README.md ("紧凑检索" section) inside otherwise-English docs. Please keep docs English (or confirm mixed language is intentional).
• Commit 4's message mentions fixing the read-only SQL keyword scan, but that landed on main via #110 — this commit doesn't touch that code.

Minor

• PiSessionUnitMeta.kind was widened to string; a template-literal type would keep some checking.
• Low-level error messages still say "Pi" for OMP files (Malformed Pi session header in ...).
• parseLines assumes the OMP title record is always at line 0; a mid-file title record would fail entry validation. Worth a comment documenting the invariant (or skipping type === 'title' in the entry loop).

Verified locally

Built the PR CLI and ran the related suites: 107 pass / 0 fail. On a real ~1 GB index, bare --search returned 68.8 KB / 140 messages; with --compact --limit 8 --context 0 --snippet-length 240 it returned 5.1 KB / 8 hits. The flags work as advertised.

One observation from real usage data (a follow-up thought, not a request for this PR): in indexed agent history, --query scripts outnumber --search roughly 10:1, and agents already truncate output with head -c, producing broken JSON. If bounding output matters, defaults (or a total-size guard) may deserve a separate discussion rather than opt-in flags only.

[Dustella]

Split the provider-agnostic query work into #127. This PR is now rebased onto current main and contains only the three OMP provider commits.