feat(orchestrator): pipeline profiling as a queryable subgraph

[Disentinel]

What

Emit the grafema analyze pipeline as a profile subgraph so the critical path + bottlenecks are a GRAPH QUERY, not log-grepping. Closes the derive-pack blind spot — derive packs were previously visible only as a Rule pack materialized pack=X ms=N edges=M log line; now each is a graph node.

Schema (profile: namespace, synthetic file __grafema_profile/<run_ts>)

Node type	one per	attrs
profile:run	analyze	ts, total_ms
profile:phase	phase (resolve/derive/…)	phase
profile:stage	resolver-cmd / derive-pack	wall_ms, edges_produced, nodes_produced, kind, phase, order
METRIC	one measure	value, unit — reuses the existing METRIC node + OBSERVES edge (same as per-file parse_ms)

Edges: profile:phase --PART_OF--> profile:run; profile:stage --PART_OF--> profile:phase; profile:stage --PRECEDES--> profile:stage (the route); METRIC --OBSERVES--> stage/phase.

On by default (cheap — a few hundred nodes). Disable with GRAFEMA_PROFILE_SUBGRAPH=0. Profile nodes are profile:-namespaced and committed after the analyze summary counts, so code queries and existing count assertions are unaffected.

Implementation

• profiler.rs — retain events in memory (ProfileEvent) + iso_now().
• analyzer.rs — profile_subgraph_to_wire() builds the WireNodes/edges (+ unit tests).
• main.rs — build_profile_stages() from the event stream; commit the subgraph at end of analyze via commit_batch.
• scripts/profile-graph.mjs — critical path (longest PRECEDES chain by summed wall_ms) + jams + dead stages, read from the graph.
• _ai/profile-subgraph.md + SKILL.md — schema + documented Datalog queries + the engine binding semantics.

Query (real output from a self-analyze on grafema-dev)

716928 nodes / 1532179 edges, 0 SIGSEGV, 46 stages.

--- critical path (longest PRECEDES chain by summed wall_ms) --- phase=derive sum=242.7s
   36.8s @stdlib/js_local_refs   edges=230
   22.8s @stdlib/rust_calls      edges=2404
   18.3s @stdlib/haskell_local_refs edges=89628
   12.8s @stdlib/js_runtime_globals_edges edges=15243
   ...
--- top jams ---
   36.8s [derive] @stdlib/js_local_refs   25.4s [resolve] js_resolve   22.8s [derive] @stdlib/rust_calls
--- dead stages (wall_ms>=1s AND edges=0) — the REG-1128 class ---
   5.8s @stdlib/shape_verifier   3.5s @stdlib/rust_imports   3.3s @stdlib/js_property_access_ns
   3.0s @stdlib/java_imports   2.7s @stdlib/go_imports   1.3s @stdlib/go_context

Pure-Datalog dead-stages (CI-gate candidate, verified on the real graph → 14 zero-edge packs):

dead(S, Name) :- node(S, "profile:stage"), attr(S, "edges_produced", "0"), attr(S, "name", Name).

Note: the derive-query engine matches metadata attrs by constant value only — it does not enumerate a metadata value into an unbound var. So numeric measures (wall_ms) are read via getNode in the helper; the edges_produced=0 dead-stage query is pure Datalog.

🤖 Generated with Claude Code