Windows: quarantined experimental-leg tests (process/path/bridge gaps after #179/#180)

[fujibee]

After the Git Bash compat (#179) and sqlite CRLF (#180) fixes landed, the windows-latest experimental leg has a residual set of failures that depend on POSIX process/path semantics agmsg does not yet support natively on Windows. To let the experimental leg report green (instead of perpetually red / 12-min-timeout cancelled), these are quarantined with skip_on_windows in the bats suite rather than deleted. Un-skip each as it is fixed.

Quarantined tests

Watcher process management (ps/pgrep/pkill + pidfiles under Git Bash):

• delivery stop: kills watchers and emits stop directive
• watch.sh restricts subscription to active_name when 4th arg is given
• delivery set turn: kills only the target project's watcher, leaves other projects'
• delivery set off: kills only the target project's watcher, leaves other projects'

Codex bridge identity resolution (returns "no matching identity" on windows-latest; likely project-path normalization in the bridge resolve path):

• codex-bridge: resolve-only prints the selected identity
• codex-bridge: resolve-only rejects ambiguous identities
• codex-bridge: rejects unsupported app-server endpoints
• codex-bridge: connects to ws://host:port app-server endpoints
• codex-bridge: refuses when the same identity already has a live bridge

Codex bridge launch / symlink path (#160 canonical-path matching + nohup background launch):

• session-start.sh for codex matches rollout cwd via a symlinked project path (#160)

commandWindows not written on Windows (the codex Stop entry's commandWindows comes back empty for a plain project path even though delivery set succeeds — distinct from the escaping fixed in #175/#134):

• delivery set turn (codex): Stop entry carries commandWindows wrapping Git Bash

Not quarantined

The test-side CRLF assertion fragility (counts / json_valid probes comparing "1\r" to "1") is fixed directly by routing test assertions through a CR-stripping sqlite_mem helper, not skipped.

Refs #125, #134, #160.

[simonlpaige]

The skip_on_windows quarantine makes sense for now, but the Codex bridge identity resolution failures (returning "no matching identity") are likely caused by path normalization differences in bridge_resolve_path() — specifically how Git Bash translates /c/Users/... vs C:\Users\... when comparing against stored project paths in SQLite.

Check src/bridge.sh around the resolve_path function: if it uses readlink -f or realpath, those have inconsistent behavior under Git Bash (MSYS2's realpath sometimes returns mixed separators). A targeted fix would normalize paths with cygpath -m before comparison, or use sqlite3's REPLACE() to collapse backslashes before matching.

For the watcher process management failures: ps under Git Bash doesn't reliably list MSYS2-origin processes with the same arguments as Linux ps. Use pgrep -f with explicit pattern anchoring, or fall back to tasklist + findstr when uname -s contains MINGW or MSYS.

The symlink path issue in session-start.sh (#160) is likely $(pwd -P) vs $(pwd -L) — on Windows junctions, pwd -P resolves the physical path while pwd -L keeps the symlink. The fix is to consistently use -P or -L depending on whether you want canonical or logical paths in the rollout cwd check.