Consolidate radio/task-pm/task-reviewer to canonical files (#122 Phase 1; precedes #163-#169)

[martin-conur]

Closes #170

Phase 1 of epic #122 — lands before the radio-reliability bundle #163–#169 so each of those tickets edits one file instead of seven.

What changed

• bin/radio — the 7 byte-identical loadout copies collapse into one canonical root file (sentinels stripped; no dispatch needed since it's identical everywhere). Every loadout install.sh now links $SCRIPT_DIR/../bin/radio.
• bin/task-pm — the root dispatcher is replaced by the shared body. Impl comes from lib/detect-impl.sh (--impl / AW_IMPL / workflow-doc auto-detect); the agent prefix (claude / kiro) branches only the final launch line (claude "/pm" vs kiro-cli chat --agent pm). 7 copies deleted.
• bin/task-reviewer — the claude×4 body is folded into the root file; kiro-* impls still route to kiro-gh/bin/task-reviewer until kiro-gh/bin/task-reviewer: migrate to SPEC_IDENTIFIER + AW_LOADOUT to match claude-* (post-#144) #146. AW_LOADOUT (tracker-aware behavior, task-reviewer: spec-issue cross-check is GitHub-only — breaks claude-jira / claude-notion / claude-local #144) now comes from impl detection instead of the file's own path. 4 copies deleted.
• tools/check-drift.sh — retired radio|body, task-pm-claude|body, task-pm-kiro|body, task-reviewer-claude|body; 17 groups remain, green.
• Tests — helpers point at the canonical binaries; task_pm.bats / task_reviewer.bats pin loadouts via AW_IMPL=<impl> run … (the workflow-doc auto-detect routing tests are unchanged and still cover detection); install_symlinks.bats asserts the radio symlink resolves to the root copy.
• task-pm --help / task-reviewer --help now work outside a configured repo (per the fresh-install acceptance criterion; previously the dispatchers errored on impl detection before help could print).

Acceptance criteria

• Exactly one radio, one task-pm; task-reviewer = 1 canonical + kiro-gh's. Note: the issue said "+ 3 kiro", but only kiro-gh ever shipped a task-reviewer (kiro-local/kiro-notion parity is kiro-gh/bin/task-reviewer: migrate to SPEC_IDENTIFIER + AW_LOADOUT to match claude-* (post-#144) #146, listed as a non-goal) — so 1 + 1 is the correct landing state.
• ./run_tests.sh: 805/805 green. tools/check-drift.sh: green (17 groups).
• Fresh ./install.sh all into a scratch $HOME: radio / task-pm / task-reviewer symlinks all resolve to the root canonicals and --help cleanly.
• task-pm behavior is byte-equal to today on both agents except the launch line (body shared, branch is the final case "$AGENT").

🤖 Generated with Claude Code

[martin-conur]

Review — PR #171 (spec: #170, epic #122 Phase 1)

Verdict: no blockers. Five nits, all optional — none should hold up the merge.

Spec compliance

Checked against #170's acceptance criteria: ✅ one radio, one task-pm in the tree; task-reviewer = 1 canonical + kiro-gh; ✅ drift manifest entries retired (radio|body, task-pm-*|body, task-reviewer-claude|body); ✅ suite green (805/805 per PR body, spot-verified paths); ✅ installers relinked. The "1 canonical + 3 kiro" → "1 canonical + 1 kiro" deviation is correctly argued in the PR body — only kiro-gh ever shipped a reviewer; kiro-local/notion parity is #146's problem. Accepted.

Verification performed

• Swept the whole tree for references to the 17 deleted per-loadout paths: zero stale references (installers, tests, docs, drift manifest, dispatchers).
• All 7 installers' ln -sf sources resolve to existing files in the new tree (../bin/radio ×7, ../bin/task-pm ×7, ../bin/task-reviewer ×5 where applicable).
• Old kiro task-pm behavior (env exports + exec kiro-cli chat --agent pm) is preserved byte-for-byte in the shared body's kiro branch.
• shellcheck -x clean on the three canonical files (CI runs shellcheck, so this matters).
• Checked the copied bodies against main: the triple gh pr view calls and double git show-ref in task-reviewer, and the double rev-parse in task-pm, are preexisting on main — consolidation copied them verbatim, so not flagged as regressions here.

Nits (optional, in rough priority order)

1. bin/task-pm:57 — the case "$AGENT" in claude|kiro has no *) fallback. Unreachable today (detect-impl validates the impl list), but if a third agent family ever lands, task-pm renames the tab, mutates env, and exits 0 without launching anything. One line: *) echo "Error: no PM launch line for agent '$AGENT'" >&2; exit 1 ;;
2. bin/task-reviewer:113 — the kiro-unavailable error still links closed Extend task-reviewer to claude-jira, claude-notion, claude-local + README/gh-workflow docs #136 while the comment two lines up correctly says kiro-gh/bin/task-reviewer: migrate to SPEC_IDENTIFIER + AW_LOADOUT to match claude-* (post-#144) #146 tracks parity. Inherited from the old dispatcher, but now user-visible next to a contradicting comment. Point it at kiro-gh/bin/task-reviewer: migrate to SPEC_IDENTIFIER + AW_LOADOUT to match claude-* (post-#144) #146.
3. bin/task-reviewer — [[ "$AW_LOADOUT" == "claude-gh" ]] is tested in 5 places. Deriving TRACKER="${AW_LOADOUT#*-}" once and branching on gh would make kiro-gh/bin/task-reviewer: migrate to SPEC_IDENTIFIER + AW_LOADOUT to match claude-* (post-#144) #146 (kiro-gh folds into this body, also a gh tracker) and kiro-jira loadout (only missing combo in the 2×4 grid) #92 (kiro-jira) nearly free instead of a 5-site hand-edit.
4. tests/install_symlinks.bats:43 — the "symlink resolves to canonical root bin/radio" assertion runs only for claude-gh, and claude-jira/claude-notion/kiro-local/kiro-notion installers aren't exercised at all. The exact regression this PR guards against (an installer pointing at a deleted per-loadout path) would ship undetected for 4 of 7 loadouts; looping the existing test over all seven installers is ~10 lines.
5. */install.sh:19-23 — the now byte-identical 3-line symlink stanza across 7 installers is a new unguarded duplication class (the retired drift entries guarded the old one). A single install-symlinks|symlink-stanza|<7 files> manifest entry in tools/check-drift.sh — mechanism already supports it — closes the gap. Alternatively fold into Epic: Loadout duplication consolidation (radio + lib-pull + prompt templates) #122 Phase 2.

Nits 4+5 could also be deferred into a small follow-up issue rather than blocking this PR — worktree churn on a −5.8k-line diff is the riskier option.

Review process note: diff reviewed via 8-angle candidate pass + verification against main; 3 efficiency candidates were discarded as preexisting-on-main rather than introduced.

[martin-conur]

Re-review — commit 09075a7

All five nits verified fixed:

1. ✅ task-pm agent case now has a loud-failing *) fallback.
2. ✅ Stale Extend task-reviewer to claude-jira, claude-notion, claude-local + README/gh-workflow docs #136 link → kiro-gh/bin/task-reviewer: migrate to SPEC_IDENTIFIER + AW_LOADOUT to match claude-* (post-#144) #146.
3. ✅ TRACKER="${IMPL##*-}" replaces all five claude-gh string checks. Also verified the removed AW_LOADOUT had no surviving references (would have been an unbound-variable crash under set -u) — the launch env uses the separate TASK_FORCE_LOADOUT, so removal is safe.
4. ✅ New test sweeps all 7 installers against a clean ~/.local/bin, including the kiro-local/kiro-notion no-reviewer assertion.
5. ✅ install-shared-symlinks sentinel region drift-guarded across all 7 installers (18 groups green); loadout-specific reviewer/board lines correctly left outside the region.

Independently verified: targeted bats suites green locally, check-drift.sh green, shellcheck -x clean, and all 4 CI checks pass on the head commit. LGTM — ready to merge.