chore(bench): fixtures + analysis/transport benchmarks + baseline

[gandarfh]

Slice 2.5 foundation: measure before optimizing (the ADR-010 principle).

• gen_fixtures.py: deterministic medium.md (466L/10 blocks) + large.md (5226L/50 blocks/~190 refs); refs point only at blocks above (real scope resolution).
• bench_analysis.ml: in-process analysis cost (scan/diagnostics/semantic/completion/hover), no transport.
• lsp_roundtrip.py: extended to run the fixtures E2E (didChange->diagnostics, semanticTokens/full) + a --check budget gate against the ADR-010 hard-fail thresholds.
• BASELINE.md: numbers + the conclusions they drive.
• CI: Linux-only budget gate in the ocaml job (hard-fail thresholds, not the noise-prone p95 targets).

Key baseline finding: the OCaml analysis is sub-millisecond on the large fixture (didChange->diagnostics 1.9ms p95). The one hot spot is semanticTokens/full on large at ~70ms p95 -- encoding/transport bound (of_blocks itself is 0.83ms), exactly what semantic tokens delta targets. The planned LSP memo/cache optimization would tune work already 100x under budget; the data redirects effort to semantic tokens delta and the frontend full-doc scanners (measured separately in httui-desktop).