fix(parser): detect shebang language from the parsed byte snapshot (#746)

A file saved while it is being indexed could be silently under-indexed:
parse_bytes hashed and parsed one byte snapshot, but for extension-less
shebang scripts detect_language re-read the file from disk. A save
(truncate+rewrite) racing that probe made a complete snapshot parse to
zero nodes, stored under the snapshot's final file hash, leaving the
graph missing the file while it looked up to date.

detect_language now accepts the pre-read source bytes and sniffs the
shebang from them; parse_bytes passes its snapshot through, so every
parse decision and the stored hash derive from the same read. Path-only
callers (change filters) keep the on-disk probe.

Regression tests simulate the race by patching the on-disk probe and
the parse-stage read, covering both the extension-less script wipe and
the stored-hash-of-parsed-snapshot invariant on the incremental path.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FmPUotoqsNWWt3b41FMjwv