fix(hooks): run every handler in the scope the dispatcher resolved (#752)

[SaulMoro]

Part of #752 (item 5).

Summary

 hookDispatchCli(event, stdin)
   chdir(payload cwd)                      ← can fail: deleted worktree
   localConfig = resolveConfigForDir(payload cwd)
   handlers = filterHandlersForConfig(registry, localConfig)
-  createDispatcher({ handlers })
+  createDispatcher({ handlers, localConfig })
-    handler.execute(stdin, tool)
+    handler.execute(stdin, tool, localConfig)

 dashboard-report  teamCorrectionKeywords
-  autoDetectInit()          ← process cwd
+  loadTeamConfig(localConfig.repo.localPath)
 votes-sync
-  autoDetectInit()          ← process cwd
+  localConfig
 webhook-dispatch
-  loadWebhookConfig()       ← autoDetectInit(), process cwd
+  loadWebhookConfig(localConfig)

With a payload cwd that no longer exists, the dispatcher resolves the user scope. The three handlers above then read the config of whichever directory the host started the hook in:

payload cwd: deleted worktree   → dispatcher: user scope
process cwd: project A          → keywords, votes, webhook: project A

The contribute hint is not affected: it already resolves from the payload cwd (#747). loadWebhookConfig() with no argument keeps its CLI behaviour (teamai webhook list / test).

Second commit: the background pass starts again when the payload cwd is gone. Found while verifying the first. The detached child that runs the background handlers was spawned in the payload cwd. On macOS and Linux a directory that no longer exists fails the spawn, the error is swallowed (child.on('error', () => {})), and no background handler runs: no session-start pull, no webhook, no update check.

 spawnPlainDetached(command, args, stdin, cwd)
-  spawn(…, { cwd })                              ← ENOENT, swallowed
+  spawn(…, { cwd: exists(cwd) ? cwd : tmpdir })  ← as the WMI launch on Windows

It needs the first commit. The child resolves its scope from the payload, not from the directory it starts in, but before this PR its handlers read the process cwd. Started from the host's directory, they would have landed in project A. The temp dir is also safer than inheriting the host's directory for the session-start pull, which still reads the process cwd (item 2 of #752).

Evidence

Red → green, hook-dispatch-scope.test.ts (real hookDispatchCli, sandbox HOME, both passes):

handlers follow the scope the dispatcher resolved, not the directory the hook process runs in (#752)
  user scope (member tester) · project A (member alice, keyword "rehazlo", webhook)
  payload cwd deleted, process.chdir(project A)
  prompt "rehazlo", then Stop with a transcript that declares a recalled doc
                                  before          after
  prompt_submit.correction        true            false
  ~/.teamai/votes/                [alice.yaml]    [tester.yaml]
  webhook requests                1               0

Real CLI (dist/index.js hook-dispatch), same setup, one sandbox per agent. Each scope's team has its own webhook (/a, /user) on a local receiver. Builds: main (dad371c), the first commit (cb1d0ff), and this PR:

build	agent	keywords	votes	webhooks received
main	claude / codex / codebuddy / opencode	correction: true (A's keyword)	alice.yaml	none
cb1d0ff	claude / codex / codebuddy / opencode	correction: false	tester.yaml	none
this PR	claude / codex / codebuddy / opencode	correction: false	tester.yaml	1 × /user

"none" in the first two rows is the background pass never starting. A control run with the payload cwd in project A fires A's webhook once (checked on both commits of this PR).

Session start with a deleted cwd, hook process in project A: with this PR the background pull runs (it installs ~/.teamai/bin/teamai, which cb1d0ff never gets to). Project A's data directory holds the same files before and after, on both builds.

Red → green, hook-dispatch-cli.test.ts:

starts the background pass from the temp dir when the payload cwd no longer exists
  spawn cwd   before: <payload cwd, deleted>   after: os.tmpdir()

Also: hook-dispatch.test.ts checks that the dispatcher hands every handler, foreground and background, the scope it was created with. The keyword tests in hook-handlers.test.ts now read a real teamai.yaml from the passed scope instead of mocking autoDetectInit.

Test plan

• npx tsc --noEmit, npx vitest run (4341 passed, 1 skipped), npm run test:e2e (228 passed, 26 skipped), npm run build
• Real CLI, git provider, Claude, Codex, CodeBuddy, OpenCode: the table above. Only the agent axis is covered: these handlers do not depend on the provider, and nothing here pushes (votes go through updateReports, which fails offline in the sandbox either way).

Merge Danger

Door: two-way. No data format or file layout changes.

Blast Radius: hooks

HookHandler.execute takes a third parameter. Every handler in the registry is updated. Handlers other than these three ignore it.

Sessions whose cwd was deleted now run their background handlers on macOS and Linux, as they already did on Windows. That includes a session-start pull of the scope the dispatcher resolved (the user scope).

[github-actions]

Findings

• [P1 blocking] src/hook-dispatch-cli.ts:110 assumes os.tmpdir() is an existing, scope-neutral directory. It honors TMPDIR/TMP/TEMP, which may be missing or located inside another TeamAI project. Because the session-start pull still detects scope from the child’s process cwd, a deleted-worktree hook can still fail to spawn or pull/deploy the unrelated project instead of the dispatcher-resolved user scope.
• [P1 blocking] The PR description’s Test Plan covers real-CLI verification only with the git provider and explicitly omits gitlab and github. The trusted AGENTS.md requires real-CLI verification across git, gitlab, and github, so the required end-to-end record is incomplete.