fix(pampa): detect parse errors without the per-lex tree-sitter logger (bd-b7eb7)

[cscheid]

Summary

The qmd reader (crates/pampa/src/readers/qmd.rs::read) attached a tree-sitter set_logger callback on every parse. When any logger is set, tree-sitter formats a debug string with snprintf for every lex action — and on macOS snprintf locks the global locale (localeconv_l). Under multithreaded project renders (rayon, ~16 workers) that locale lock serializes all parser threads.

How it was found

Profiled a large healthy project (401 docs / 8.2 MB, no errors) — the case we most want fast. A fresh multithreaded render spent ~75% of samples in __ulock_wait2/__ulock_wake/os_unfair_lock, call chain:

ts_lexer__advance → snprintf → __vfprintf → localeconv_l → os_unfair_lock

16 threads bought only ~1.2× wall-clock; sys time ~37 s. It is not the allocator — swapping mimalloc both as #[global_allocator] and via tree_sitter::set_allocator changed nothing. The locale-lock culprit was found with macOS sample (which symbolicates system libraries, unlike samply's sidecar).

Fix

Parse the happy path with no logger, and detect whether tree-sitter entered error-correction mode via its own public signal — the parse_with_options progress callback's ParseState::has_error — combined with the final tree's Node::has_error (which backstops an error in the last few ops before parse end). The logger is attached only for the diagnostic re-parse, which runs only when the document has an error. Encapsulated in MarkdownTree::had_parse_error. The previous, ineffective TreeSitterLogObserverFast (its cheap callback didn't help — the snprintf is inside tree-sitter, before the callback) is removed.

Why has_error is reliable here: tree-sitter-qmd declares no conflicts:, so the parser is deterministic LR and GLR multiple-stack speculation only occurs during error recovery — has_error therefore means a genuine error, not a speculative branch that will be pruned. This invariant is now documented in grammar.js and parser.rs.

Results (large healthy project, release-perf)

metric	before	after
16-thread render	8.04 s	3.71 s (2.2×)
16-thread sys	36.83 s	1.37 s (27×)
pass-1 @ 16 threads	2935 ms	426 ms (scales ~6.4× on 16 cores)
single-thread render	9.71 s	6.0 s (1.6×)

Correctness

• Error-corpus snapshot tests (json / text / ariadne / locations) pass.
• The real 565-file qmd-plans corpus reports the same 140 error files — exact parity with the previous log-based detection.
• Full cargo xtask verify (Rust workspace + WASM/hub build + hub tests) passes.

Notes

• Also documents the profiling lesson (use macOS sample for system-lib frames; a Rust #[global_allocator] doesn't cover a vendored C library's libc malloc) in claude-notes/instructions/performance-profiling.md.
• Full investigation: claude-notes/research/2026-05-31-q2-render-perf-qmd-plans.md (lands separately with the website experiment).

🤖 Generated with Claude Code

[cscheid]

Superseded by #247 — same commits, rebranched to perf/tree-sitter-error-logging.