v0.6.0

Changed

• Extraction is substantially faster. Configurations, packages, and the
  manifest's opening calls now run concurrently, every nix subprocess
  drawing from one shared gate sized two below the core count (clamped
  2–8, overridable with FLAKE_EXPLORER_NIX_JOBS on memory-tight
  hosts). Option chunks are batched into one nix process each, so a
  batch pays the flake-load and module-system setup (~580ms on a real
  configuration) once instead of per chunk; a failing batch splits
  rather than degrading, so no chunk loses detail because of who it was
  batched with. And each extraction now records how the option tree
  split, so the next extraction of the same configuration — after an
  upgrade or a refresh — replays that plan instead of rediscovering the
  poisoned paths by binary search. Measured on a real 15,436-option
  NixOS configuration: whole-flake --all dropped from ~146s to ~94s,
  single-configuration extraction a further 21% with 39% less peak
  memory, and a re-extraction with a recorded plan runs 1.82x faster
  than a cold one with 47% fewer nix evaluations. Extracted output is
  byte-identical throughout.
• The extraction cache invalidates far less often. Extraction code
  moved into its own crate (crates/extract), and the cache
  fingerprint now covers exactly that crate, its resolved dependency
  closure, and the nix version — previously it hashed every source file,
  so a UI-only patch release threw away every user's cached extractions.
  The boundary is structural rather than a curated list: code outside
  the crate cannot participate in extraction, and a write-site test
  guards against new persistence sneaking into the root crate. This
  release itself moves the fingerprint, so cached extractions
  re-extract once on upgrade.
• Extracted data is now deterministic. Persisted option order and
  warning order previously depended on which worker finished first, so
  the same flake could produce different bytes on machines with
  different core counts — options now sort by source location and
  warnings by the option path they concern, and a test holds the whole
  data directory byte-identical across repeated --all runs.
• CI is roughly six times faster (~825s to ~136s): coverage gets a
  dependency layer that actually matches, checks build in the dev
  profile without debug info, thin LTO, a parallel job graph, and a
  cargo cache keyed on flake.lock instead of a rustc it never used.
  The two coverage suites (crate and SPA) now post one clearly labelled
  PR comment instead of two identical-looking ones.

Added

• Benchmarking infrastructure, and the baselines it exists to defend.
  FLAKE_EXPLORER_TIMINGS=1 makes each extraction pass report its wall
  clock on stderr without altering stdout by a byte.
  scripts/bench-extract.ts measures cold and warm extractions with
  per-phase medians, CPU, and peak RSS; scripts/bench-ab.ts runs
  interleaved, lock-guarded A/B comparisons between two binaries, and
  scripts/datadir-diff.ts verifies their outputs byte-for-byte.
  bench/BASELINES.md records what extraction costs at each measured
  commit, and bench/PHASE2-RECON.md documents where the options pass
  spends its time and why. Measurement only — none of it ships in the
  binary or touches the extraction fingerprint.

Fixed

• serve no longer re-serves a stale error when a request lands in the
  window between an extraction finishing and its bookkeeping being
  cleared: clicking refresh after a failed extraction could return the
  previous failure without extracting anything. The race required CPU
  contention to hit, which is why it appeared only as a CI flake.