fix(herdr): make the herdr adapter usable on Windows

[bingb0t5]

Herdr is the only runtime available on the Windows host these were found on (tmux, zellij, orca and cmux are all absent), so every dispatch goes through it. Three POSIX-only assumptions in the adapter each refused work outright. Two of them are fixed here; both are narrow and independently useful on any platform.

1. Socket path assumed POSIX

fm_backend_herdr_canonical_socket_path accepted only paths starting with /. Herdr on Windows reports C:\Users\...\AppData\Roaming\herdr\herdr.sock in both HERDR_SOCKET_PATH and herdr session list --json, so the launcher-identity check refused every spawn with "reports an unusable socket path". Both sides report the identical value, so accepting the drive-letter spelling is sufficient. Separators fold so the two Windows spellings compare equal, including when the directory no longer resolves.

2. jq emits CRLF

jq on that host opens stdout in text mode: printf '{"a":["x","y"]}' | jq -r '.a[]' returns x\r\ny\r\n. Command substitution strips only the trailing CR, so single-value reads are clean but multi-line reads keep an interior CR on every line but the last. That corrupted ids split out of multi-line captures and operator-facing messages. Three multi-line captures in the adapter now strip at source.

Not audited here: jq is used throughout bin/, so other scripts capturing multi-line jq output on Windows may have the same defect.

3. mkdir -m 700 cannot produce mode 700 on NTFS

The presentation lock namespace had to be owned by this user and be exactly mode 700. Git Bash mounts NTFS noacl by default, where chmod is a silent no-op and every directory reads 755, so the assertion could never hold. The namespace was permanently unobtainable, and because teardown_herdr_preflight_target resolves that lock for every herdr endpoint, teardown refused forever with no manual recovery: a finished task could not be collected, its metadata could not be cleared, and supervision stayed permanently "needed" as a result.

Ownership and the symlink refusal are unchanged. The exact-700 assertion is still required verbatim wherever the filesystem can store it; only where the filesystem provably cannot is ownership accepted as the available guarantee.

Capability is probed, not inferred: create a directory, ask for 0700, read back what was stored. A filesystem that discards the request answers for itself, so no uname or /etc/fstab parsing decides a security-relevant branch. The relaxation is also narrow in practice: on that mount Git Bash maps /tmp to the user's own profile temp directory (usertemp), so the namespace is not in a shared location and Windows access control, rather than emulated mode bits, is what actually guards it.

This also repairs test_presentation_session_lock_path_is_shared_across_homes, which fails on unmodified main on such a host.

Testing

tests/fm-backend-herdr.test.sh against this branch on the Windows host: 169 pass, 0 fail, exit 0 - the full suite runs clean to completion. Unmodified main aborts at test 20 on the CRLF fault, so most of the suite was never reached there at all.

New coverage accompanies each fix: drive-letter socket acceptance, unchanged POSIX and refusal behaviour, separator folding when the directory is gone, and the probed-capability namespace check.

Two notes for anyone else running the suite on Windows:

• It is slow, and it emits dofork: child -1 ... Resource temporarily unavailable retry noise from Git Bash partway through. That is fork-retry churn, not failure; given enough time the suite completes cleanly. A short timeout will make it look like a hang.
• ShellCheck is not installable on that host, so bin/fm-lint.sh could not be run for CI parity. Lint review is worth a close look.

A fourth Windows fault is not addressed here: herdr pane get reports no foreground_cwd field at all, so the worktree-entry poll cannot succeed. A candidate fix exists but is deliberately held back until a real spawn is observed to succeed.

🤖 Generated with Claude Code

https://claude.ai/code/session_01F7L1m56txHs7hrNcTeq7AP

[kunchenguid]

Speaking as Kun's firstmate: closing this as stale. It has been waiting on a contributor update for 14+ days with no author push or comment. Reopen if you want to pick it back up.