Commit f828151d5f

Phase 7: parse diagnostics with source locations, and rayon parallelism

Parse diagnostics:
- The parser always returns a document — malformed constructs degrade rather than
  crash. The gap was that they degraded silently, and the worst cases are severe: an
  unterminated #+BEGIN_SRC reads the rest of the file as block content, and an
  unterminated drawer does the same but renders to nothing, so one missing line deletes
  most of a page from a build that reports success.
- parse now returns Document::diagnostics, each carrying a 1-based source line, for
  unterminated blocks, unterminated drawers (including :PROPERTIES:), and stray #+END_
  with nothing open. The build prints file:line: message; --strict makes them, and
  unresolved links, a non-zero exit.
- Line numbers are threaded as an absolute offset through every nested parse, so a
  block inside a list item inside a section reports its real file line. There is a test
  for exactly that, because reconstructed and re-indented nested slices are where an
  off-by-N hides. Threading the offset surfaced a shadowing bug: parse_list already had
  a local `base` for the indent column, so the line offset was silently shadowed;
  renamed to base_indent.
- The 179-file corpus produces zero diagnostics.

Parallelism (rayon over PARSE, RESOLVE and RENDER/EMIT):
- 179 files: 0.23s -> 0.07s. 1,790 files: 3.98s -> 0.82s (4.9x on 12 cores).
  RAYON_NUM_THREADS=1 reproduces the old 3.98s exactly, so the gain is parallelism and
  not incidental change, and output is byte-identical to the sequential build across
  the whole corpus.
- PARSE is pure over one file's bytes and RESOLVE only reads the shared symbol table,
  which is what makes them safe to parallelize; INDEX stays sequential.
- Parallelism must not be observable in the result. par_iter().collect() preserves
  order so the bytes are unaffected, but the report is the fragile half: pushing to
  rendered/skipped from inside the parallel pass would order them by thread scheduling,
  a non-deterministic report over a deterministic site. The parallel pass returns only
  what was written and the report is assembled sequentially after. The new determinism
  test was verified by reintroducing that bug and watching it fail.

Measured but deliberately not fixed: the build is O(n^2) because the nav bar lists every
page. At 1,790 pages each page carries 1,799 nav links and the output is 284 MB against
5.5 MB for 179 pages — 52x the bytes for 10x the input. Parallelism moves that constant
without fixing it. Which pages belong in a nav is a product decision this project has
not made, so it is recorded in the README rather than guessed at.

Layout: unified · split