Fix #1034: split transcript basenames on backslash too, so the scan cache works on Windows

[fdaviddpt]

Closes #1034

Diagnosis and the proposed expansion are @RidoMr's, from the issue: thank you for the clear repro and the before/after numbers.

What was wrong

Claude Code on Windows sends transcript_path as a native C:\Users\...\<session-id>.jsonl path. After #829 decodes the JSON escapes, TRANSCRIPT holds single backslashes and no / at all. post-tool-hook.sh derived basenames with ${TRANSCRIPT##*/}, so the whole path survived, failed the [A-Za-z0-9._-] sanitiser and came out empty. The #913 tmp/transcript-scan.<id> sidecar was then never read or written, and every PostToolUse call outside the save cooldown cold-rescanned the whole transcript. Nothing was logged.

Change

• scripts/post-tool-hook.sh: all three basename sites now use ${TRANSCRIPT##*[/\\]} -- the scan-cache name, the SESSION_ID fallback (reached when a payload carries transcript_path without session_id), and the text of the Basename SESSION_ID sanitiser (#620) can empty to "" with no non-empty gate, turning a loud refusal into a silent full-transcript save #633 "failed the session id sanitiser" log line. The ./../leading-dash/charset sanitiser is untouched, so nothing new can reach the sidecar path or the save-session argv.
• Silent-failure guard: when the cache name comes out empty for a non-empty TRANSCRIPT, the hook logs one WARNING: transcript-scan cache disabled for "<path>" ... line, once per transcript. The marker tmp/transcript-scan-unnamed holds the last transcript warned about; it is read with the read builtin and written with a builtin redirect, so the guard adds no fork on any path and nothing at all when the name is usable (Windows 25H2: every hook fork leaks a kernel Token object until reboot (win32k CForegroundLaunch bug) — fork tax is a persistent leak, not just latency #1011/post-tool-hook creates ~67 processes per no-save tool call #1029 budgets).
• tests/test_locale_ranges_695.py: four allowlist line numbers shifted by the new comment and guard lines (no new range added).
• changelog.d/1034.fixed.md.

Tests (tests/test_post_tool_backslash_transcript_basename_1034.py)

Runs the real hook with a stdin payload, through resolve_bash() so it runs on the Windows leg (unlike the #913 module, which is skipped there).

• Native backslash transcript_path writes tmp/transcript-scan.<id> after one call -- Windows only, because str(Path) there is the real native shape. Skipped elsewhere with a reason: on POSIX a backslash is a filename byte, so it would be a different file, not the Windows path.
• Forward-slash transcript_path, same fixture -- positive control, every platform.
• SESSION_ID fallback (no session_id in the payload) checked through tmp/capture-alive.d/<id>: backslash case on Windows, forward-slash control everywhere.
• POSIX-only: a file whose name literally contains backslashes now splits at the last one. This pins the one behaviour change the fix makes on POSIX; it is not Windows evidence. It is also what let me watch the module go red locally.
• Guard: an unusable name (bad+name.jsonl) logs the warning exactly once across two calls (must fire); a good name logs nothing (must not fire).

Observed vs reasoned

• Observed (macOS, bash 5.3 and /bin/bash 3.2.57): without the hook change the module fails 3 (both POSIX literal-backslash cases and the must-fire guard) and passes 3 (the forward-slash controls and the must-not-fire case); with it, 6 pass and the 2 Windows-only cases skip, under both bashes. The ${p##*[/\\]} expansion gives the expected basename for POSIX, drive-letter forward-slash, native backslash and mixed-separator paths on bash 3.2.57 and 5.3. Adjacent suites (post-tool-hook: ~1.2 s per tool call on Windows (0.33.0) - whole-transcript wc -l grows with the session; lib-slug spawns cygpath twice per call #913, Windows 25H2: every hook fork leaks a kernel Token object until reboot (win32k CForegroundLaunch bug) — fork tax is a persistent leak, not just latency #1011, safe_eval drops every bridge variable containing 'I' under tr_TR locale — no save ever completes #695, Hook stdin extractors never decode JSON escapes: Codex on Windows yields a doubled-backslash PROJECT_DIR and the wrong session slug #829, The first-match stdin extractor now feeds cwd on all four hooks, and #344's top-level-wins pin only ever covered source #447, PostToolUse still cannot save on Codex: the session id is taken from the transcript basename, which #459 stopped controlling #468, post-tool-hook.sh: STDIN_SESSION_ID guard gap lets a flag-shaped session_id reach save-session.sh's argv (same class as #576/#600) #610, post-tool-hook.sh basename-derived SESSION_ID still reaches argv unguarded (leading-dash gap closed on two of three routes) #620, Basename SESSION_ID sanitiser (#620) can empty to "" with no non-empty gate, turning a loud refusal into a silent full-transcript save #633, post-tool-hook.sh _pt_scan_lines: quadratic-cost scan hangs the hook on real transcripts (v0.42.5 release-audit, round 1, BLOCKING) #1017, post-tool-hook.sh: marker mtime stops refreshing when SESSION_ID is empty (v0.42.5 release-audit, round 1, non-blocking) #1019, The per-tool-call hot path lost its byte-identity pin and the replacement can be walked around #330, resolve-paths.sh tests REMEMBER_HOOK_CWD for directory existence before the Windows path normalization that would make it resolvable #448, Release tree: inline the sourced shell libraries into each hook so the directory validator can follow every hook (clears COMMAND_SCRIPT_NOT_FOLLOWED) #900 compile/strip) pass locally: 189 passed, 4 skipped.
• Observed by the reporter (Windows 11, MSYS bash 5.2.37), not by me: the expansion result on native paths, and the sidecar appearing and hook time dropping after the same change on a real machine.
• To be observed in this PR's CI: the native-backslash sidecar and SESSION_ID cases on windows-latest. Until that leg is green, the Windows claim here is reasoned.
• Reasoned: the SESSION_ID fallback route matters in production only for payloads without session_id (Claude Code's normal payload carries both). Whether Codex on Windows also sends backslash paths is not checked; if it does, the same change covers it.

🤖 Generated with Claude Code

[AI-generated]