test: add action_sequence grader tests (microsoft#110)

Related to microsoft#105

Proactive test cases for the action_sequence grader. Written from the
spec while Linus implements.

Covers: exact_match, in_order_match, any_order_match modes, score
calculations, error cases, and factory integration.

**Test breakdown (31 test cases):**
- Constructor validation: 6 tests (missing actions, empty actions,
invalid mode, valid modes)
- exact_match mode: 6 tests (pass, extra step, missing step, wrong
order, empty actual, single tool)
- in_order_match mode: 5 tests (exact seq, extras interspersed, wrong
order, missing step, subset order)
- any_order_match mode: 5 tests (reordered, extras, missing, duplicate
expected, insufficient duplicates)
- Score calculations: 4 tests (perfect, partial, zero, details
verification + f1 check)
- Edge cases: 3 tests (nil session, nil tools_used, duration recording)
- Factory integration: 2 tests (Create success, Create invalid params
error)

**Note:** These tests depend on the implementation in the companion PR.
Merge the implementation PR first.

Author: spboyer

.ai-team/agents/linus/history.md

@@ -21,4 +21,8 @@
 - **Mock engine (`execution.NewMockEngine`)** — Returns deterministic responses. Useful for testing the full pipeline without a real Copilot SDK connection. Spec YAML with `executor: mock` triggers it.
 - **Fixture isolation** — The runner resolves `--context-dir` relative to CWD and task glob patterns relative to the spec file directory. Tests use `t.TempDir()` for full isolation.
 - **Shared workspace hazard** — Multiple agents may work concurrently in this repo. Always verify `git branch --show-current` before committing; stash operations can carry changes across branches.
+- **Nil session pattern (#105 review fix):** Graders that depend on `SessionDigest` must return a graceful zero-score `GraderResults` (not an error) when `session == nil`. The runner treats grader errors differently from zero-score results — errors may abort the run while zero-scores are recorded as failed validations. Follow `behavior_grader.go` as the canonical pattern. Fixed in `action_sequence_grader.go` per PR #110 review from Rusty.
+📌 Team update (2026-02-11): Grader nil-session error handling contract — all graders depending on SessionDigest must return zero-score GraderResults with nil error when session is nil, not (nil, error). — decided by Rusty
+
+📌 Team update (2026-02-11): PR #103 azd extension packaging approved and merged. Registry URLs need update to spboyer/waza for production. — decided by Rusty
 <!-- Append new learnings below. -->

.ai-team/agents/rusty/history.md

@@ -9,4 +9,5 @@
 ## Learnings
 
 <!-- Append new learnings below. Each entry is something lasting about the project. -->
+📌 Team update (2026-02-11): Run command tests must reset package-level flag vars (contextDir, outputPath, verbose) at top of each test body to prevent state leakage. — decided by Linus
 <!-- Append new learnings below. -->

.ai-team/decisions.md

@@ -8,12 +8,36 @@
 **What:** All issues must follow: feature branch → commit → push → PR with `Closes #N` → @copilot review → address feedback → merge. No direct commits to main.
 **Why:** User directive
 
-### 2026-02-09: Sensei reference repo
+### 2026-02-09: Sensei reference repo for E2 implementation
 **By:** Squad (Coordinator)
-**What:** E2 sensei engine must adopt patterns from https://github.com/spboyer/sensei. Scoring: Low/Medium/Medium-High/High. Ralph loop: READ→SCORE→CHECK→SCAFFOLD→IMPROVE→VERIFY→TOKENS→SUMMARY→PROMPT→REPEAT. Token management: count/check/suggest/compare with .token-limits.json config.
+**What:** The sensei engine (E2: #32-38) must adopt functionality from https://github.com/spboyer/sensei. Key patterns to port to Go:
+
+**Scoring Algorithm:** Low (description < 150 chars OR no triggers) → Medium (description >= 150 chars AND trigger keywords) → Medium-High (USE FOR + DO NOT USE FOR) → High (Medium-High + routing clarity). Rule-based checks: name validation, description length, trigger/anti-trigger detection, routing clarity, MCP integration.
+
+**Ralph Loop:** READ → SCORE → CHECK → SCAFFOLD → IMPROVE FRONTMATTER → IMPROVE TESTS → VERIFY → CHECK TOKENS → SUMMARY → PROMPT → REPEAT (max 5 iterations, target: Medium-High).
+
+**Token Management:** count/check/suggest/compare with .token-limits.json config. SKILL.md: 500 soft / 5000 hard. References: 2000 each. --strict exits 1 on limit breach.
 **Why:** User directive — spboyer/sensei is the reference implementation
 
 ### 2026-02-09: Monitor human engineer comments
 **By:** Shayne Boyer (via Copilot)
-**What:** Periodically check for comments from Charles and Richard on open issues. Follow up on responses.
+**What:** Periodically check for comments from Charles (@chlowell) and Richard (@richardpark-msft) on open issues. Follow up on responses.
 **Why:** User directive
+
+### 2026-02-09: Run command test patterns
+**By:** Linus (Backend Dev)
+**Related:** #24
+**What:** Each test that calls `cmd.Execute()` must reset package-level flag vars (`contextDir`, `outputPath`, `verbose`) at the top of the test body to prevent state leakage. If new flags are added to `newRunCommand()`, the reset block must be updated accordingly.
+**Why:** Package-level Cobra flag bindings persist across test cases in the same process.
+
+### 2026-02-11: PR #103 — azd extension packaging approved
+**By:** Rusty (Lead)
+**Related:** E7 (#62), PR #103
+**What:** Wallace Breza's azd extension PR approved and merged. Adds `extension.yaml`, build scripts, `registry.json`, `version.txt`. Purely additive — no changes to existing Go code or CI. Follow-up: update registry URLs to `spboyer/waza` for production; add trailing newlines (POSIX nit); consider CI for automated extension builds.
+**Why:** E7 (AZD Extension) P2 roadmap item — establishes foundation without disruption.
+
+### 2026-02-11: Grader nil-session error handling contract
+**By:** Rusty (Lead)
+**Related:** PR #110
+**What:** All graders depending on `SessionDigest` must return a zero-score `GraderResults` with `nil` error when session is nil — not `(nil, error)`. Matches `behavior_grader.go` pattern (lines 56-62). Returning an error would abort the entire run instead of gracefully recording a zero score.
+**Why:** Runner distinguishes grader errors (may abort) from zero-score results (contribute to scoring). Affects `action_sequence_grader.go` and future session-dependent graders.

.ai-team/decisions/inbox/coordinator-sensei-reference.md

@@ -0,0 +1,11 @@
+# Session Log: 2026-02-11 — PR Reviews
+
+**Requested by:** Shayne Boyer
+
+## Events
+
+1. **Rusty reviewed PR #110** (action_sequence grader): requested changes — nil session bug must return zero-score result not error; alignment nit on constant naming.
+
+2. **Rusty reviewed PR #103** (Wallace's azd extension): approved — clean, purely additive, aligns with E7 roadmap. Registry URLs need update before production.
+
+3. **Linus fixed PR #110 review feedback**: added graceful nil session handling (zero-score return), updated test assertions, aligned constants.

.ai-team/decisions/inbox/copilot-directive-2026-02-09-git-workflow.md

@@ -0,0 +1,11 @@
+## Orchestration Entry
+
+| Field | Value |
+|-------|-------|
+| **Agent** | Linus (Backend Dev) |
+| **Routed because** | Reviewer rejection lockout: Basher authored PR #110, Rusty requested changes → Linus assigned revision |
+| **Mode** | background |
+| **Files authorized** | action_sequence_grader.go, action_sequence_grader_test.go, outcome.go |
+| **Files produced** | Updated 3 files, pushed to squad/105-action-sequence-tests |
+| **Outcome** | ✅ Fixed nil session handling (graceful zero-score), test assertions, and constant alignment per Rusty's review |
+| **PR** | #110 |

.ai-team/decisions/inbox/copilot-directive-2026-02-09-sensei-ref.md

@@ -0,0 +1,11 @@
+## Orchestration Entry
+
+| Field | Value |
+|-------|-------|
+| **Agent** | Rusty (Lead) |
+| **Routed because** | Code review of PR #103 (Wallace's azd extension) |
+| **Mode** | background |
+| **Files authorized** | PR #103 diff (read-only) |
+| **Files produced** | PR review comment on #103 (approved), decisions/inbox/rusty-pr103-review.md |
+| **Outcome** | ✅ Approved. Clean, purely additive, aligns with E7. Non-blocking notes about registry URLs and trailing newlines. |
+| **PR** | #103 |