test(adapters): collapse the sanitize_path suite, and fix the bypass it found

[hsusul]

Stacked on #9, #10, #12 — merge those first.

Problem

Two hand-written assertion lists covered one 25-line pure function with 88 assert_eq! lines. 69 of them arrived one per commit — 12% of this repository's entire history is single-assert commits against sanitize_path. That history is unreviewable, and enumerating slash/dot permutations by hand only ever tests inputs someone thought of.

Change

sanitize_path_table — all 87 unique cases preserved verbatim as data, grouped by what each demonstrates. I extracted them mechanically rather than by hand, with a parser that correctly handles Rust raw strings ending in a backslash (r"src\app\"), and cross-checked the extracted count against the assert_eq! count in each block so nothing was silently dropped. Adding a case is now a row, not a commit.

sanitize_path_invariants — six properties that must hold for all input: no .. segment survives, no empty or . segments, never absolute, no control or zero-width characters, no drive prefix, and idempotence.

The property test found a real bug

Not a test problem — a genuine bypass in sanitize_path:

sanitize_path("A\x7f:")                    ->  "A:"              (expected "")
sanitize_path("C\u{200b}:/Users/dev/x.rs") ->  "C:/Users/dev/x.rs"

The drive-letter check tested the raw segment, before control and zero-width characters were filtered out. Slipping any such character between the letter and the colon made the segment longer than 2 bytes, so the length check missed it — then the character was stripped a moment later and the drive prefix survived into the output.

Windows drive-prefix neutralization is a stated boundary-hardening control, so this was a real hole in it, not a theoretical one. Fixed by filtering first and testing the filtered segment. Four regression cases are in the table alongside the property that caught it.

This is the argument for the change in one example: 88 hand-written cases never found it; the property found it on the first run.

Dependency

proptest as a dev-dependency only, default-features = false — drops the fork/timeout machinery and its rusty-fork + wait-timeout deps, keeping the graph small for the egress audit. egress-check.sh still passes.

Verification

Gate	Result
cargo fmt --all -- --check	pass
cargo clippy … -D warnings	pass
cargo test --locked -p lore-core -p lore-ipc --all-targets	346 pass, 4 ignored
cargo test -p lore-core --test no_network_in_archive	pass
./scripts/egress-check.sh	pass
./scripts/toolchain-pin-check.sh	pass
sanitize_path_invariants × 5 runs	stable over fresh case sets each run