fix(probe): exclude volatile env pseudo-vars from probe key (#201)

[emmanuelm41]

Closes #201.

Problem

On the Windows e2e arm (#82) the compiler-probe memo missed on the warm build, re-running the probe — probe_runs = 1 where c-hello / c-depinfo assert max_probe_runs = 0:

{"name": "max_probe_runs", "expected": "<= 0", "actual": "1", "passed": false}

Everything else on that phase passed (compile local_hit, compiler_runs: 0, diff byte-identical, verify ok) — only the probe memo failed to hit. Not a #196 regression.

Root cause

The probe key (src/probe/cache.rs::probe_key) mixes in env_fingerprint(), which hashes the entire process environment. On Windows std::env::vars() surfaces cmd.exe's hidden =-prefixed pseudo-variables:

• =C:, =D:, … — the per-drive current working directory
• =ExitCode — the previous child process's exit status

These shift between the cold and warm make invocations, so the probe key differed across builds and the warm load() looked up a {key}.json that cold never wrote → miss → re-probe.

This is Windows-only because Unix variable names cannot contain =. It breaks only the probe memo (not the compile cache) because the main cache key (cache_key.rs) hashes only a curated env subset (RUSTFLAGS, CARGO_ENCODED_RUSTFLAGS, CARGO_CFG_*) — which stays stable — exactly matching the observed "compile hits, probe re-runs".

Fix

Filter =-prefixed names out of the env fingerprint. Refactored the hashing into a pure fingerprint_env core (+ is_volatile_env_name predicate) so the filtering is unit-testable without mutating the real environment. No-op on Unix.

Tests

Added 3 regression tests (fingerprint_env_ignores_windows_hidden_pseudo_vars, fingerprint_env_still_reflects_real_vars, is_volatile_env_name_flags_only_equals_prefixed). cargo fmt, cargo clippy --all-targets, and cargo test -p kache (626 passed) all green locally.

Validation: the Windows e2e arm (non-blocking, self-hosted kunobi-windows runner) on this PR's CI run should now show c-hello / c-depinfo warm with probe_runs = 0.

[emmanuelm41]

Superseded by #209, which consolidates the remaining Windows e2e fixes (this PR's changes are included there as a dedicated commit) so the whole arm can be driven to green in one CI run. Closing in favor of #209.