refactor(workflows): compress pending_review_gate to emit + script-poll pattern

[PolyphonyRequiem]

Summary

Compresses the two pending_review_gate nodes in github-pr.yaml and ado-pr.yaml from blocking human_gate nodes into an emit + script-poll pattern, following the gate-compression-pattern ADR (docs/decisions/gate-compression-pattern.md).

Gates compressed: 2
Both pending_review_gate nodes were category-1 (deterministic-outcome): the workflow was pausing to wait for a human to take an action on the PR (merge, approve, comment, close) that Poll-PrStateDelta.ps1 can detect directly. They stay compressed.

Gates staying as human gates: 17
poll_error_gate, revise_cap_gate, remediation_cap_gate, stuck_review_gate, pr_pre_merge_gate, merge_failed_gate and others — all genuine judgment calls, config errors, or infrastructure failures. They are unchanged.

---

Per-file changes

github-pr.yaml (v2.4.8 → v2.5.0)

• Added poll_timeout_seconds (default 86400 s / 24 h) and poll_interval_seconds (default 30 s) as optional workflow inputs
• Added notifications: block (polyphony.github_pr namespace) with types pr_review_required and pr_ci_attention_required
• Replaced three nodes (pending_poll_counter + pending_review_gate_policy_router + pending_review_gate) with:
  • notify_pr_pending — emits pr_review_required domain signal once when gate opens
  • poll_pr_state_delta — calls Poll-PrStateDelta.ps1; blocks until reaction or timeout
  • notify_pr_ci_attention — emits pr_ci_attention_required on CI FAILURE (Q3 Option C)
• Added notify_pr_review_resolved before already_merged_emitter (emits disposition:resolved to close the pending CTA)
• Updated stuck_review_gate prompt (removed pending_poll_counter template refs; timeout-based description)
• Rewired stuck_review_reset → poll_pr_state_delta
• Rewired pr_pre_merge_gate.defer → poll_pr_state_delta

ado-pr.yaml (v2.4.8 → v2.5.0)

Same pattern as GitHub with ADO-specific details:

• poll_pr_state_delta uses -Platform ado; coords parsed from PrUrl by the script
• notify_pr_pending payload includes organization and project
• CI attention threshold uses 'failed' (ADO build policy status string)
• ADO vote mapping documented in node comments (+5/+10 → approved, -10/-5 → changes_requested)

---

Script: Poll-PrStateDelta.ps1

Added Liszt's implementation (was untracked). One small contract gap patched:

WatermarkPath was mandatory in Liszt's implementation but optional in Wagner's plan §3.2. Made optional: when omitted, the path is auto-derived from PR coordinates (platform-owner-repo-number). This allows the YAML to omit -WatermarkPath entirely.

Reaction-kind contract note

Liszt's actual reaction_kind enum differs slightly from Wagner's plan. The YAML routing uses Liszt's actual values:

Liszt's value	Meaning
initial_observation	First call, watermark just established — YAML routes back to re-poll
pr_merged / pr_closed	Terminal PR states
new_review_approved / new_review_changes_requested / new_review_commented	Review reactions
new_commit / new_comment	Commit/comment activity
ci_status_changed	CI status flip (with delta_details.new for green/red routing)
timeout	Budget elapsed — escalates to stuck_review_gate_policy_router

on_error: → poll_error_gate is deferred to AB#3257 (conductor RFC Phase 2 — on_error: in routes not yet shipped in v0.1.18).

Pester tests for the script (Poll-PrStateDelta.Tests.ps1) are included and were AST-clean per Liszt. Full test execution is follow-up work.

---

ADRs added

• docs/decisions/gate-compression-pattern.md — the pattern this PR implements (Bach)
• docs/decisions/domain-signal-envelope.md — dependency (Bach)
• docs/decisions/polyphony-verb-error-boundary.md — dependency (Bach)
• docs/north-star.md — high-level vision doc (Bach)

---

Open follow-ups

• AB#3257: Add on_error: to: poll_error_gate to poll_pr_state_delta once conductor RFC Phase 2 ships
• Post upstream PR Closed-loop step 2: wire PlanObserver into state next-ready #213 cherry-pick (27006af): bulk-rename type: notification → type: emit and notification: → emit: in both files
• Run Poll-PrStateDelta.Tests.ps1 in CI
• Thread poll_timeout_seconds / poll_interval_seconds through feature-pr.yaml and implement-merge-group.yaml callers

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com

[PolyphonyRequiem]

📊 Pester verification — 5 pre-existing bugs caught + patched

Ran scripts/Poll-PrStateDelta.Tests.ps1 locally on this branch and found 5 bugs that pre-date this PR:

1. Script — Invoke-AssertCli accessed nonexistent \$Result.Exe under StrictMode → PropertyNotFoundException collapsed 401/404/429 to exit code 5. Fixed by removing the redundant guard (ExitCode -eq -1 + 'not found on PATH' regex already suffices).

2. Test — Build-GhStub PR-core glob 'api repos/*/pulls/* --jq *' also matched /pulls/{n}/reviews because * is greedy. Tightened to --jq {* (jq filter for the PR-core call starts with {).

3. Script — Find-GitHubDeltas returns a typed List[hashtable] through the pipeline: empty list → \, single item → bare hashtable. \.Count was unreliable in both cases. Wrapped call-site in @() to always produce a plain array.

4. Script — Select-HighestPrecedenceDelta declared [System.Collections.Generic.List[hashtable]]\. StrictMode threw PropertyNotFoundException when a plain hashtable (single-delta) was passed. Removed the type constraint.

5. Test — ConvertFrom-Json auto-promotes ISO-8601 strings to DateTime objects; Should -Match was receiving a DateTime's ToString() output. Fixed by normalising to string before asserting. Also bumped one test's TimeoutSeconds from 10→30 so CI-stub subprocess latency doesn't cause a spurious timeout.

Fixed in commit b37d85f. Pester is now 30/30 green (1 intentional skip for the 60s+ hang test).

— Liszt