spec(t27f): raise SpecGraph capacity to 1024 nodes and 4096 edges so every spec in the corpus is shown #10230

SpecGraph holds at most 192 nodes and 254 edges (#10180), so about 2 percent of the corpus is still refused with ERR_TOO_MANY_NODES: the biggest specs (gf16 436 nodes, t27core 678) are exactly the ones a reader most wants to see. The owner wants big specs shown, not refused.

Measured on the Railway lab over all 2225 specs/**/*.t27 files at master 538df07, through the exporter and then the derive pass (#10198), on a scratch build with caps of 2048 nodes and 8192 edges so that nothing is refused: nodes p50 20, p90 65, p95 93, p98 191, p99 444, p999 720, max 775; edges p50 18, p90 118, p95 211, p98 371, p99 531, p999 1573, max 3194. All 2225 specs validate. The largest by nodes is specs/igla/coder/eval.t27, the largest by edges is specs/tools/t27o/rust_convert.t27.

This issue raises MAX_NODES to 1024 and MAX_EDGES to 4096 (headroom of 1.3x and 1.28x over the measured maximum), moves the edge endpoint arrays and the NO_INDEX sentinel from u8 to u16 where the old limit forced it, and keeps refusing, never truncating, anything past the new caps. Epic: gHashTag/trinity#1648. Predecessors: #10180 (192/254), #10162 (exporter), #10198 (derive).

Boundary

• specs/t27f/spec_graph.t27
• specs/t27f/spec_graph_export.t27
• specs/t27f/spec_graph_derive.t27
• specs/t27f/spec_graph_fold.t27
• .claude/skills/t27f/SKILL.md

spec_graph_derive.t27 and spec_graph_fold.t27 (landed in #10228 while this was in review) are in the Boundary only for the minimum a wider index forces: their as u8 casts at with_edge, stale comments, and (derive) two capacity tests that hard-coded 192 and 254. spec_graph_view.t27 and spec_graph_outline.t27 are NOT touched: the view reads indices as u32, so it adapts as it stands.

User Scenarios

• Given any spec in specs/ when the exporter and the derive pass read it then validate_graph answers GRAPH_OK and the graph keeps every node and edge.
• Given a spec with 1025 nodes or 4097 edges when the exporter scans it then the verdict is ERR_TOO_MANY_NODES or ERR_TOO_MANY_EDGES and no truncated graph is returned.
• Given a graph with a node past index 255 when an edge names it then the edge is stored with its exact endpoint and the graph validates.

Requirements

• FR-001: MAX_NODES MUST be 1024 and MAX_EDGES MUST be 4096; every fixed array in SpecGraph and in the exporter's name-span table MUST have that length.
• FR-002: edge_from and edge_to MUST be u16 arrays, with_edge MUST take u16 endpoints, and NO_INDEX MUST be 65535; an invariant MUST state that both caps stay below NO_INDEX and keep at least 25 percent headroom over the measured maximum.
• FR-003: A graph exactly at each cap MUST be valid and one past it MUST be refused, tested in spec_graph, from real source text in spec_graph_export, and through the derive pass in spec_graph_derive.
• FR-004: The measured distribution and its method MUST be written next to the constants in spec_graph.t27, and SKILL.md MUST state the new limits.
• FR-005: The PR MUST state the wasm size and the per-call time of the exporter on the largest specs, before and after.

Success Criteria

• t27c test-report reports 0 vacuous and 0 failing tests for spec_graph, spec_graph_export, spec_graph_derive, spec_graph_outline, spec_graph_view and spec_graph_fold.
• t27b test --check (qemu-aarch64) reports 0 failed tests for the same six specs.
• Exporter plus derive validate 2225 of 2225 specs in specs/ on the lab.
• The pull request touches 0 files outside the Boundary section, adds 0 lines of non-t27 code, and leaves spec_graph_view.t27 unchanged.