ACE-045: per-process model cache + single SQL parse + shared guard indices

[ashwin-agami]

Summary

Behaviour-preserving perf hardening (audit finding P2, feature F12-runtime-scalability). The semantic model is parsed/indexed far more than needed per query. Two independent wins, same answers:

• Slice 1 — GuardContext (both paths). The _model_safety guard battery re-parsed the SQL 6× and rebuilt each model index per guard. A GuardContext now parses once + builds each index once, threaded through the guards.
• Slice 2 — get_cached_org (long-lived server). The parent loaded the whole model 2–3× per query and re-loaded it every query. It's now cached per process, keyed (org_id, datasource, model_version), served warm across queries + users.

Changes

• semantic_model/runtime.py — GuardContext + build_guard_context; optional ctx= on the 6 guards + _preflight_select/_check_aggregation_semantics. ctx=None keeps the standalone (cli.py) path byte-identical.
• execute_sql.py — build the context once in _model_safety, thread it; rebuild after an auto-rewrite (SQL changes).
• tools.py — get_cached_org + _current_org_ctx/_current_org_id; route _resolve_units/_resolve_receipt/tool_get_datasource_schema. Bypass the cache when model_version is None (never serve a stale model).
• mcp_http.py — set _current_org_ctx per request in handle_mcp from the resolved org (mirrors _actor_ctx). Tenant-safe by construction: the cache key is org-scoped via the OrgResolver seam — N=1 today, per-request under a multi-tenant resolver, no fork, no F14 dependency.
• tests/ — test_guard_context.py, test_org_cache.py, conftest.py (isolates the module-global cache).
• Synced the bundled lib/ mirror.

Test plan / acceptance

• Slice 1: parse=1 + each-index-built=1 counter; verdict-parity with/without ctx across 6 SQL cases; unparseable→allow.
• Slice 2: 1 load/query + 0 on repeat; reload after a version bump; two orgs on the same datasource get separate models (no cross-tenant); default org_id → local.
• Full gate green: 1413+ tests, ruff + format + gitleaks + lib-drift all pass (uv run dev.py check).

Out of scope (per spec)

Cross-query caching on the subprocess/file path (needs ACE-028); the multi-tenant resolver (agami-hosted F5) — this only consumes the OrgResolver port.

Spec: ACE-045

[Copilot]

Pull request overview

This PR implements ACE-045 performance hardening by (1) introducing a per-invocation GuardContext to avoid repeated SQL parsing and repeated semantic-model index builds across the safety-guard battery, and (2) adding a per-process semantic-model cache (get_cached_org) scoped by (org_id, datasource, model_version) so long-lived servers reuse warm models across queries while remaining tenant-safe via a request-scoped ContextVar.

Changes:

• Add GuardContext / build_guard_context and thread ctx= through guard functions to reuse a single parsed SQL tree and shared indices.
• Add per-process org/model caching in tools.py, plus request-scoped org-id propagation in the HTTP MCP transport to keep caching org-scoped.
• Add tests covering guard-context reuse/parity and org-cache warm hits, version invalidation, and org scoping.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
tests/test_org_cache.py	Adds Slice 2 tests validating per-process org/model caching behavior and org scoping.
tests/test_guard_context.py	Adds Slice 1 tests ensuring single-parse/single-index-build and parity with/without ctx.
tests/conftest.py	Adds a global autouse fixture to isolate module-global org cache state between tests.
packages/agami-core/src/semantic_model/runtime.py	Introduces GuardContext and updates guards to optionally reuse parsed SQL + indices.
packages/agami-core/src/execute_sql.py	Builds guard context once and threads it through the safety battery; refreshes after rewrite.
plugins/agami/lib/execute_sql.py	Syncs the same guard-context threading change into the bundled lib/ mirror.
packages/agami-core/src/tools.py	Adds (org, datasource, version) model caching and uses it in tool paths that previously reloaded models.
packages/agami-core/src/mcp_http.py	Sets _current_org_ctx per request so the per-process model cache remains tenant-scoped.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[ashwin-agami]

Thanks @copilot — all three addressed:

• build_guard_context builds indices even when sqlglot is unavailable — good catch; it now returns `None` early when `not _HAVE_SQLGLOT`, since every guard short-circuits to allow before it touches the context. Fixed centrally in `build_guard_context` so both call sites (initial + post-rewrite rebuild) benefit.
• pre_flight_check parity for unparseable SQL — real divergence in the byte-identical `.as_dict()` invariant this spec claims: with `ctx`, `ctx.tree` is `None` → "no SELECT; skipped", but the non-ctx path raised → "unparseable; skipped (…)". Unified by routing the non-ctx branch through the same `_parse_sql` helper, so both paths are now identical.
• duplicate autouse cache fixture — removed the per-file `_isolate_cache`; `conftest._reset_org_cache` already covers it (and dropped the now-unused `pytest` import).

Gate green (1417 passed). Pushed.