…rk (#465)

* fix(#464): PR-create hooks resolve PR origin repo, not session ops-fork

When the session is pinned to the ops-fork but `gh pr create --repo X`
targets a sibling repo, two PreToolUse hooks evaluated the WRONG repo:

- require-agdr-for-arch-pr.sh diffed the ops-fork working tree (showing
  framework paths .claude/migrations/, topologies/, handbooks/domain/)
  instead of the PR's actual diff, blocking even with a valid AgDR.
- validate-pr-create.sh used the ops-fork's `upstream` remote as a ticket-
  existence fallback for the sibling PR, causing false "ticket not found"
  blocks when a premium ticket number collided with a closed framework issue.

Fix:
- Add _lib-pr-repo.sh with `pr_repo_matches_cwd` helper: parses --repo from
  the command, resolves `git remote get-url origin` for the cwd, compares
  them. Handles SSH + HTTPS URL forms on macOS (BSD sed-safe patterns).
- require-agdr-for-arch-pr.sh: exit 0 early when the PR targets a repo that
  differs from the current git working tree's origin — the diff would be
  meaningless from the wrong cwd.
- validate-pr-create.sh: suppress the upstream-remote fallback when CMD_REPO
  is set and the target repo belongs to a different project lineage (matches
  neither origin nor its upstream).

Framework gating preserved:
- Framework PRs (--repo = ops-fork origin) still diff the framework tree and
  block on arch-path changes without an AgDR.
- Fork → upstream PRs (#207) still use the upstream fallback when CMD_REPO
  matches origin or upstream of the current working tree.

Tests: test_pr_hooks_cross_repo.sh — 9 cases covering A1-A4 (agdr hook) and
B1-B5 (validate hook), including both the fix scenarios and regressions.

Refs #464

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(#464): harden cross-repo guard — WARN on missing lib, B6 allow-path test

Finding 1 — silent revert on missing lib:
  require-agdr-for-arch-pr.sh: when _lib-pr-repo.sh is absent, emit a loud
  WARN naming the missing lib and that repo-resolution is degraded. The hook
  now falls through to the diff check visibly rather than silently skipping
  the guard entirely (which re-opened bug #464 invisibly).

  validate-pr-create.sh: same shape — the cross-repo guard's else branch
  (inline fallback) now emits a WARN before proceeding. The guard logic is
  preserved via the inline fallback, but the degraded state is no longer silent.

Finding 2 — test doesn't exercise the new guard allow-path:
  test_pr_hooks_cross_repo.sh B6: new test where --repo=<fork-origin>,
  ticket exists only in upstream, and the upstream remote is configured.
  CMD_REPO_LC == ORIGIN_LC → guard's allow-path fires → upstream fallback
  permitted → PASS. A deliberate bug in the guard (wrong variable in the
  comparison) makes B6 FAIL. Verified fails-before/passes-after.

  Added C1 + C2: assert the WARN is emitted when the lib is missing.

Finding 3 (optional, done cheaply):
  _lib-pr-repo.sh pr_repo_matches_cwd: when git_origin_repo returns empty
  (no origin remote), emit a WARN before returning 0 (safe default) so the
  "no origin → treat as same-repo" assumption is visible, not silent.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(#464): handle --repo=VALUE, -R VALUE, -R=VALUE forms in all three parsers (F1, F2, F3)

F1 — all four --repo/-R forms now parsed in all three parsers:
  - _lib-pr-repo.sh: pr_cmd_target_repo() extended to match
    --repo VALUE, --repo=VALUE, -R VALUE, and -R=VALUE. The
    BSD-sed alternation-capture-group approach (^|[[:space:]]) is
    replaced with the greedy `.*[[:space:]]<FLAG>` form which
    correctly extracts only the value on macOS sed -E.
  - validate-pr-create.sh: CMD_REPO extracted via pr_cmd_target_repo
    from _lib-pr-repo.sh (DRY); inline fallback updated with the
    same four-form logic for partial-checkout environments.
  - require-agdr-for-arch-pr.sh: CMD_REPO_PRESENT grep updated to
    match all four forms via `(--repo[=[:space:]]|-R[=[:space:]])`.
  New tests: A5/A6/A7 (agdr hook) and B7/B8/B9 (validate hook)
  confirm the guard fires for each form; all six tests fail when the
  parser is intentionally broken to only handle --repo VALUE.

F2 — C1 in test_pr_hooks_cross_repo.sh refactored to sandbox-copy
  pattern (same as C2). The previous C1 used mv to remove
  _lib-pr-repo.sh from $REPO_ROOT/.claude/hooks/ and relied on mv
  back after the test; a SIGKILL between the two mv calls would
  permanently remove the lib from the live hooks directory. C1 now
  copies the hook + its deps into a temp sandbox, removes the lib
  from the sandbox copy only, and never touches $REPO_ROOT/.claude/.

F3 — Two stale comments in C1 ("exits 0") updated to "BLOCK / rc=2"
  to match the actual expected behavior asserted by the test.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

---------

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>