pan-wizard 3.35.0 → 3.37.0

package/docs/archive/SKILLS-ENHANCEMENT-PLAN.md

@@ -0,0 +1,230 @@
+# Skills Enhancement Plan
+
+Systematic review of all 42 shipped PAN Wizard skills against 20 modern prompting
+techniques researched from 2025-2026 best practices. Each enhancement is one
+self-contained change, applied and verified individually.
+
+**Sources:** Anthropic Prompting Best Practices, Claude Code Best Practices,
+Paxrel (10 Agent Patterns), Agentic AI Design Patterns 2026, BrightCoding
+System Prompts Guide, PromptLayer Flow Engineering, Deepset Context Engineering.
+
+---
+
+## Current State Assessment
+
+**What PAN already does well:**
+- Phase-gated execution (Plan-Act-Verify) in exec-phase, focus-auto
+- State handover protocol in pause/resume (state.md)
+- Structured JSON output enforcement in focus-scan, focus-plan
+- Guard rails (PAN_SOURCE_ROOT, NEVER DO sections in focus-auto)
+- Parallel tool directives in exec-phase (wave-based)
+- Supervisor/orchestrator pattern in plan-phase, map-codebase
+
+**What's missing or weak across multiple skills:**
+
+| # | Enhancement | Technique | Skills Affected | Priority |
+|---|-------------|-----------|----------------|----------|
+| 1 | Add Chain of Verification checklists | Self-Check Before Commit | exec-phase, quick, focus-exec | HIGH |
+| 2 | Add Error Recovery Classification | Recoverable vs Unrecoverable | exec-phase, focus-auto, focus-exec, debug | HIGH |
+| 3 | Add Motivation to Constraints | Explain the "Why" | focus-auto (NEVER DO), focus-exec (9 rules) | HIGH |
+| 4 | Add Investigate-Before-Answering directive | Anti-hallucination | verify-phase, assumptions, audit-deployment | HIGH |
+| 5 | Add Tool Selection Heuristics | Priority-ordered routing | focus-scan, focus-design, map-codebase | MEDIUM |
+| 6 | Soften aggressive directive language | Claude 4.6 calibration | focus-auto, focus-exec, focus-scan | MEDIUM |
+| 7 | Add Dry-Run verification step | Simulate Before Execute | exec-phase, quick, focus-exec | MEDIUM |
+| 8 | Add Context Window Management rules | Retention policies | focus-auto, focus-design, exec-phase | MEDIUM |
+| 9 | Add Reflexion Loop to verification | Generate-Critique-Revise | verify-phase, plan-phase (plan-checker) | MEDIUM |
+| 10 | Standardize XML structural tags | XML-First Architecture | Inconsistent across skills — some use XML, some don't | LOW |
+| 11 | Add Anti-Overengineering directive | Scope containment | exec-phase, quick, focus-exec | LOW |
+| 12 | Add Thinking Tags guidance | Mandatory reasoning | debug, focus-design, assumptions | LOW |
+
+---
+
+## Enhancement Details
+
+### Enhancement 1: Chain of Verification Checklists
+**Technique:** Self-Check Before Commit
+**Target skills:** exec-phase, quick, focus-exec
+**What to add:** A numbered verification checklist that the agent MUST run through
+before committing any code change. Currently these skills say "test after changes"
+but don't have a structured self-check protocol.
+**Checklist template:**
+```
+Before committing, verify ALL:
+1. All modified files were read before editing
+2. Tests pass (run test suite, record count)
+3. No unintended files in git diff
+4. Commit message accurately describes only verified changes
+5. No TODO/FIXME introduced without tracking
+If any check fails: fix and re-verify. Do not commit until all pass.
+```
+
+### Enhancement 2: Error Recovery Classification
+**Technique:** Recoverable vs Unrecoverable errors
+**Target skills:** exec-phase, focus-auto, focus-exec, debug
+**What to add:** Explicit classification of error types with recovery protocols.
+Currently focus-auto has "revert fast" (5 min limit) but doesn't distinguish
+error types. exec-phase delegates to workflows but has no error taxonomy.
+**Classification template:**
+```
+RECOVERABLE (retry with analysis, max 3 attempts):
+- Test failure after code change → read error, fix, re-test
+- File not found → search for moved/renamed file
+- Merge conflict → attempt auto-resolution, then checkpoint
+
+UNRECOVERABLE (halt, report to user):
+- Persistent test regression after revert
+- Permission/auth errors on critical paths
+- State corruption (malformed JSON in planning files)
+- Build system failure unrelated to current changes
+```
+
+### Enhancement 3: Motivation-Enriched Constraints
+**Technique:** Explain the "Why" behind rules
+**Target skills:** focus-auto (NEVER DO section), focus-exec (9 Behavioral Rules)
+**What to add:** A "Why" annotation to each constraint. Currently the NEVER DO
+list in focus-auto has 11 bare rules with no reasoning. Adding motivation helps
+the model generalize to edge cases.
+**Example transformation:**
+```
+Before: "- Skip the baseline test capture (Phase 1)"
+After:  "- Skip the baseline test capture (Phase 1) — without a baseline, the
+          regression circuit breaker cannot detect test count decreases"
+```
+
+### Enhancement 4: Investigate-Before-Answering Directive
+**Technique:** Anti-hallucination through evidence requirement
+**Target skills:** verify-phase, assumptions, audit-deployment
+**What to add:** Explicit directive to read referenced files before making claims.
+verify-phase currently says "validate built features" but doesn't mandate reading
+the actual implementation before judging. assumptions should read the codebase
+before surfacing assumptions.
+**Directive:**
+```
+NEVER claim a feature works or doesn't work without reading the implementation.
+Before each verification judgment:
+1. Read the source file(s) that implement the feature
+2. Read the test file(s) that cover it
+3. Run the tests
+4. Only THEN state your assessment with file:line evidence
+```
+
+### Enhancement 5: Tool Selection Heuristics
+**Technique:** Priority-ordered tool routing
+**Target skills:** focus-scan, focus-design, map-codebase
+**What to add:** Explicit tool priority ordering. focus-scan currently says
+"Grep for patterns" but doesn't prioritize which tools to try first. focus-design
+has 10 phases with many tool options but no routing guidance.
+**Heuristic:**
+```
+Tool priority (use the simplest sufficient tool):
+1. Grep/Glob — for finding specific patterns or files
+2. Read — for examining known files
+3. Bash — only for commands that dedicated tools can't do
+4. Agent (subagent) — only for exploration spanning >5 file reads
+5. WebSearch/WebFetch — only when local sources are exhausted
+```
+
+### Enhancement 6: Soften Aggressive Directive Language
+**Technique:** Claude 4.6 calibration
+**Target skills:** focus-auto, focus-exec, focus-scan
+**What to change:** Replace over-emphasized "CRITICAL", "MUST", "ABSOLUTELY"
+language with normal direct instructions. Claude 4.6 overtriggers on aggressive
+directives — they can cause overreaction or over-caution.
+**Example transformations:**
+- "You MUST ask the user" → "Ask the user"
+- "Do NOT proceed past this point" → "Wait for the user's reply before proceeding"
+- "ALWAYS EXCLUDE these directories" → "Exclude these directories from scanning"
+Note: Keep NEVER DO / ALWAYS DO sections — those are structural, not tonal.
+
+### Enhancement 7: Dry-Run Verification Step
+**Technique:** Simulate Before Execute
+**Target skills:** exec-phase, quick, focus-exec
+**What to add:** For STANDARD and FULL tier items, state the intended changes
+before implementing them. Currently exec-phase goes straight to execution.
+**Template:**
+```
+For M+ items, before implementing:
+1. State: "I will modify [files], adding [what], to achieve [goal]"
+2. Identify risks: "This could break [X] if [Y]"
+3. Then implement
+```
+
+### Enhancement 8: Context Window Management Rules
+**Technique:** Retention policies for long sessions
+**Target skills:** focus-auto, focus-design, exec-phase
+**What to add:** Explicit guidance on what to keep vs. summarize vs. discard
+as context accumulates. focus-auto can run 10+ cycles; focus-design has 10
+phases. Both risk context rot without management rules.
+**Template:**
+```
+Context management across cycles/phases:
+KEEP: Current task goals, test baselines, error states, file paths being modified
+SUMMARIZE: Previous cycle results (1-line per cycle), completed sub-tasks
+DISCARD: Raw tool output from previous cycles, superseded scan results
+```
+
+### Enhancement 9: Reflexion Loop in Verification
+**Technique:** Generate-Critique-Revise
+**Target skills:** verify-phase, plan-phase (plan-checker interaction)
+**What to add:** After initial verification judgment, explicitly re-evaluate
+against the phase goals. Currently verify-phase checks features against UAT
+but doesn't self-critique its assessment.
+**Template:**
+```
+After initial assessment:
+1. Score each requirement: PASS / PARTIAL / FAIL
+2. For PARTIAL/FAIL: state what specifically is missing
+3. Re-read the requirement and the code — did you miss anything?
+4. Revise scores if needed
+5. Only report final scores after this review cycle
+```
+
+### Enhancement 10: Standardize XML Structural Tags — SKIPPED
+**Technique:** XML-First Architecture
+**Target skills:** Many — inconsistent usage across skills
+**Decision:** Keep the current split. Short workflow-delegating skills (30-80 lines)
+use XML tags (`<objective>`, `<process>`, `<context>`). Long focus skills (200-1000+
+lines) use markdown headings for navigability. This is a sensible natural split —
+forcing XML on 1000-line files adds noise, forcing markdown on 40-line files loses
+structure. No change needed.
+
+### Enhancement 11: Anti-Overengineering Directive
+**Technique:** Scope containment
+**Target skills:** exec-phase, quick, focus-exec
+**What to add:** Explicit instruction not to expand scope beyond the task.
+focus-auto already has "Follow the Plan — no scope creep" but exec-phase and
+quick lack this. Claude 4.6 tends to over-engineer.
+**Directive:**
+```
+Implement exactly what was planned. Do not:
+- Add features not in the plan
+- Refactor surrounding code
+- Add comments to unchanged code
+- Create abstractions for one-time operations
+- Add error handling for impossible scenarios
+```
+
+### Enhancement 12: Thinking Tags Guidance
+**Technique:** Mandatory reasoning before complex actions
+**Target skills:** debug, focus-design, assumptions
+**What to add:** Encourage explicit reasoning before tool use in investigative
+skills. debug already uses scientific method but could benefit from structured
+observe-think-act loops.
+
+---
+
+## Execution Order
+
+Apply enhancements in priority order, one at a time:
+1. Chain of Verification (exec-phase, quick, focus-exec)
+2. Error Recovery Classification (exec-phase, focus-auto, focus-exec)
+3. Motivation-Enriched Constraints (focus-auto, focus-exec)
+4. Investigate-Before-Answering (verify-phase, assumptions, audit-deployment)
+5. Tool Selection Heuristics (focus-scan, focus-design, map-codebase)
+6. Soften Aggressive Language (focus-auto, focus-exec, focus-scan)
+7. Dry-Run Verification (exec-phase, quick, focus-exec)
+8. Context Window Management (focus-auto, focus-design, exec-phase)
+9. Reflexion Loop (verify-phase, plan-phase)
+10. XML Standardization (all skills — needs design decision)
+11. Anti-Overengineering (exec-phase, quick, focus-exec)
+12. Thinking Tags (debug, focus-design, assumptions)

package/docs/audits/AUDIT-2026-08-e2e-ROADMAP.md

@@ -0,0 +1,117 @@
+# E2E Audit 2026-08 — Remediation Roadmap (Part 2)
+
+**Companion to:** [`AUDIT-2026-08-e2e.md`](AUDIT-2026-08-e2e.md) (findings) · **Created:** 2026-08-05
+
+> Point-in-time planning doc; counts describe the audit run and are exempt from the live-counts policy.
+
+## Status — Round-5 FINAL: all 137 original findings resolved; 2-item fix-pass ledger + a 14-item consistency tail (2026-08-06, uncommitted working tree)
+
+> **Round 5** verified the Round-4 consolidation (see the Round 5 section of [`AUDIT-2026-08-e2e-VERIFICATION.md`](AUDIT-2026-08-e2e-VERIFICATION.md)): **17 findings — 1 High fixed immediately, 2 record errors corrected in place, 14 documentation/lint-hardening items recorded and deliberately deferred.** The High was in `.gitignore`: the ADR-0026 `!experiments/**` carve-out is the last matching pattern for everything under `experiments/`, so it undid **every** blanket rule above it — `.env`, `.env.*`, `node_modules/`, `*.log`, `.DS_Store`, `.vscode-test/`, `.codex/` were all committable in a public repo. Every safety-critical rule is now re-asserted after the negation and verified in both directions (secrets and junk ignored, including nested; harvested telemetry still committable). No product defect above Low is open. The Round-4 status below stands otherwise unchanged.
+
+## Status — Round-4: all 137 original findings resolved; 2-item fix-pass ledger (2026-08-06, uncommitted working tree)
+
+> **Round 4** (see the Round 4 section of [`AUDIT-2026-08-e2e-VERIFICATION.md`](AUDIT-2026-08-e2e-VERIFICATION.md)) closed the Round-3 ledger and is entirely **working-tree work** — there is no commit to cite for it yet. **7 of the 8 items open at the end of Round 3 are closed:** **M6** (the prescribed `hooks/` migration anchor landed at `bin/install.js:849`, first clean verdict after three PARTIALs), **N20** (new `tests/statusline-guard.test.cjs` with all three prescribed tests), **N16** (outside-sandbox assertion in the Codex `--global` test), **N24** (platform-aware win32 bridge test), **N25** + **N27** (the single-slot idempotency marker replaced by a bounded per-transcript FIFO signature set, preserved through the L40 prune), and **N12** (the model-version phrasing pass). **Still open (2), both in the cost/trace-logger ledger row:** **N26** (PARTIAL — same-type sibling rows can still collapse in the pre-existing last-row dedup, because the per-invocation discriminator that distinguishes their signatures is never written into the row) and new **N29** (Low — byte-identical legitimate spawns stay suppressed on runtimes that supply no per-invocation field). They want one ledger-schema decision, not two patches. Round 4 also carried a **user-directed model-version-agnostic sweep** (capability phrasing instead of hardcoded model gates in shipped content and live docs; new `tests/model-version-drift.test.cjs` as the anti-drift lint; a real code fix so `detectModelCapabilities` no longer returns an all-false `unknown` — and a false "lacks 1M context" installer warning — for unrecognized Claude ids), a **correction pass** for three High false claims that sweep introduced (an invented capability probe, a false future-model promise, and an invented model gate for `--hierarchical` in five docs — no model gate exists; the constraint is the runtime), and a final consolidation after a verification pass over that correction. **Cumulative: all 137 original audit findings resolved; the open ledger holds only fix-pass items N26 and N29. No Critical ever found; no High open anywhere.** History below is preserved verbatim; do not re-inflate.
+
+## Status — Round-3 verdict: 136/137 original findings resolved; 7-item residual ledger (2026-08-05, N12 closed 2026-08-06)
+
+> *Superseded by the Round-4 block above; kept as the Round-3 record.*
+>
+> **Round-3 final verification** of the `05e8f5b` sweep (see the Round 3 section of [`AUDIT-2026-08-e2e-VERIFICATION.md`](AUDIT-2026-08-e2e-VERIFICATION.md)): **7/11 verified FIXED** (N15 win32 bridge live-verified on Windows, N17-core, N14, N10, N18, N19, N22) — the first round of this chain with **no new High regressions**. Still open: **M6** (narrowed residual: exact-basename `statusline.js` customs still clobbered/corrupted — fix by anchoring the Fix #330 migration on `hooks/statusline.js`), **N20** (the sweep's "statusline-guard coverage" claim was FALSE — the diff contains zero statusline tests; three prescribed tests in the report), **N16** (missing outside-sandbox assertion), and new **N24–N27** (win32-bridge test gap; cost/trace-logger idempotency-marker redesign: interleaved dual-registration phantom re-admission, same-agent sibling drops, L40 prune erasing the marker). **N12 was closed post-sweep on 2026-08-06** by the model-version phrasing pass: the three named spots (TROUBLESHOOTING.md's hierarchical section, `commands/pan/plan-phase.md:137`) no longer gate on a model version, and both generated SKILLS docs were regenerated via `scripts/generate-skills-docs.py` — no `Opus 4.7`/`Opus 4.8` string survives in either. That closure drops the ledger from **8 items to 7** and changes no other item's status. *(Arithmetic corrected in the Round-4 consolidation: the ledger standing at the end of Round 3 was M6, N12, N16, N20, N24, N25, N26, N27 — eight findings; the earlier "7 items to 6" counted the grouped N25/N26/N27 entry once. The per-ID enumeration in the verification report is authoritative.)* **Cumulative: 136 of the original 137 audit findings fully verified resolved; M6 alone partially open. No Critical ever found; no High open anywhere.** Suite at `05e8f5b`: **3473 / 744 / 0 fail**. History below is preserved verbatim; do not re-inflate.
+
+## Status — Round-2 re-audit of the `b1fa91f` sweep: 25/29 verified; 11 items open (2026-08-05)
+
+> **History (do not re-inflate):** the original banner falsely claimed "ALL 137 RESOLVED." The independent fix-verification re-audit ([`AUDIT-2026-08-e2e-VERIFICATION.md`](AUDIT-2026-08-e2e-VERIFICATION.md) @ `4baf609`) corrected that to **122/137 verified**, with **15 outstanding** (H7, M6, M26, M33, M34, M45, M57, M60, M61, M62, M64, M65, L34, L47 + L36) and **14 fix-pass regressions (2 High: N1 uninstall-deletes-unparseable-settings, N2 update.md local/global collapse)**.
+>
+> **The `b1fa91f` sweep** claimed all 29 (15 outstanding + N1–N14). The **Round-2 independent re-audit** (see the Round 2 section of the verification report) verified **25 of 29 FIXED** — including both High regressions with live repro across all 5 runtimes in local and global modes, and the previously-missing M30/M32 CLI exit-code tests (N5). **Still open (4):** M6 (root cause is the untouched Fix #330 migration at `bin/install.js:840-850`, which renames non-PAN `*statusline.js` commands and defeats the new guard), N10 (verify-stubs scan-scope doc claim), N12 residual (Opus 4.7 refs in USER-GUIDE/MIGRATION), N14 (L2 guard test un-gitignored fixture + swallowed assertion). **The sweep introduced N15–N23** (9 unique; **2 High:** N15 — the M60 mode check kills the hook bridge on Windows entirely; N16 — the new Codex global-install test deletes/overwrites the developer's real `~/.agents/skills` on every suite run). Full suite at `b1fa91f`: **3466 / 742 / 0 fail** — green, but N15/N16/N20 show the suite is blind to exactly these classes. **Open ledger: M6, N10, N12, N14, N15, N16, N17, N18, N19, N20, N22** (N21/N23 were doc errors, corrected).
+
+- **Part 1 (9 High):** 8 fully verified fixed via live reproduction; **H7 PARTIAL** (Example 6's closing "Continue normally" block still shows the old renumbering flow).
+- **Part 2 (80 Medium + 48 Low):** all 7 batches below were executed, but verification upheld **14 outstanding items**: NOT_FIXED M33, M61, M62, L36; PARTIAL M6, M26, M34, M45, M57, M60, M64, M65, L34, L47.
+  - **A** parity lint + path sweep (M37/M41/M52/M54/M55) + preventive `tests/shipped-content-prefix.test.cjs`
+  - **B** dead gates & exit codes (M5/M9/M28/M30/M31/M32)
+  - **C** core correctness/robustness/security (M3/M6/M8/M10–M29/M79)
+  - **D** shipped-content drift (M1/M2/M4/M7/M34–M56/M80)
+  - **E** docs corrections (M63–M78)
+  - **F** test-coverage holes (M57/M58/M59)
+  - **G** Low-tail hygiene (L1–L48)
+- **The fix pass introduced 14 new confirmed issues (N1–N14** in the verification report**)**, including 2 High: uninstall can delete a user's unparseable `settings.json` (`bin/install.js`), and the M55 fix collapsed `update.md`'s local/global detection so global installs update as LOCAL.
+
+Full suite: **3445 tests / 733 suites / 0 failures** at `4baf609` — but N5 (promised M30/M32 exit-code tests never added), N13 (vacuous traversal assertion), and N14 (swallowed guard assertion) meant green CI under-covered the fixed gates. The remainder of this document is the original planning record.
+
+---
+
+The audit's own "highest-leverage" guidance drives the ordering: close whole *classes* systemically (with a lint that prevents recurrence) before grinding individual items.
+
+---
+
+## Batch A — Runtime-parity lint + path sweep (systemic; do first)
+
+**Why first:** this is the dominant High theme (H4/H6) and it recurs across the Medium tail. The root cause is that the installer's path-rewrite only recognizes the canonical `~/.claude/` prefix; any other prefix ships verbatim and breaks the 4 non-Claude runtimes + global installs. Fix the sources **and add a shipped-content lint** so the class can never come back.
+
+- **A0 (preventive):** new test/lint asserting shipped content (`commands/`, `agents/`, `pan-wizard-core/workflows|templates|references/`) contains only the canonical `~/.claude/` prefix — no `./.claude/`, bare `.claude/`, or `$HOME/.claude/`. This guards H4, H6, and every item below.
+- **Items:** M37 (`/pan:learn` `.claude/workflows/learn.md`), M41 (`/pan:settings` `@./.claude/…`), M52 (new-project.md `./.claude/` everywhere), M55 (update.md `$HOME/.claude/` not rewritten), M54 (global defaults path mismatch `~/.pan-wizard-core` vs `~/.pan-wizard`, stray `~/.gsd`).
+
+**Effort:** M · **Risk:** low (mechanical) · closes a class + prevents recurrence.
+
+## Batch B — Dead gates & wrong exit codes (trust-critical)
+
+**Why:** same class as H3 — commands that *claim* to gate but always exit 0, so orchestrators/CI can't rely on them. Audit every CLI path routed through `output()` (which now takes an `exitCode` arg after H3).
+
+- M9 (`campaign due` always exits 0), M30 (`verify stubs --gate` no-op on blocking findings), M32 (`learn lint` never non-zero despite docs), M28 (`dashboard --raw` discards output via raw=false), M31 (`validate health --repair` writes legacy config that disables the verifier gate), M5 (installer model-capability warning is dead code — `finishInstall` undefined `targetDir` in a swallowed catch).
+- **Add CLI-level exit-code tests** for each (the H3 gap: pure functions were tested, the exit path was not).
+
+**Effort:** M · **Risk:** medium (behavior change on exit codes — check CI/scenario consumers).
+
+## Batch C — Core-module correctness & robustness
+
+Crashes on malformed input, silent no-ops, and false "success" returns. User-facing bugs.
+
+- **Crashes / injection / containment (do first, security-adjacent):** M22 (regex injection in `requirements mark-complete` — unescaped ID → `new RegExp`), M23 (`optimize apply` writes report-supplied relative paths with no containment check — path escape), M79 (Windows backslash `..` bypasses `saveLocalPatches` traversal guard), M10 (`scaffold` without `--phase` → raw TypeError), M24 (`phases list --include-archived` ENOENT), M3 (unguarded `fs.unlinkSync` aborts whole install).
+- **Silent no-ops / false success:** M11 (`config-set` missing value reports updated:true), M27 (`state resolve-blocker` resolved:true with no match), M20 (`git commit` leaks subcommand token into files list — stages stray paths), M33 (executor blocker example omits `--text` → never recorded).
+- **Wrong results / dead code:** M12 (phase-number normalization mismatch), M13 (`cost.rates` overrides skipped for versioned model ids), M19 (frontmatter inline-array lossy on `, `), M21 (`memory optimize` agent-log consolidation dead code), M26 (`experiment stop` can never stop — pid persisted too late), M8 (`isDreamDue` contract contradiction, no consumer), M14/M15 (distill false "phantom try/catch" + "REGRESSED" labels), M16 (`experiment new` installer step always fails in installed projects), M17/M38 (counts computed from source-repo layout → wrong in host projects), M18 (`prompts_complete` stop reason unreachable), M6 (statusline clobbers Gemini/Copilot custom statuslines), M25 (review-deep ignores `--cwd`), M29 (`validate deployment` hook-path check validates nothing).
+
+**Effort:** L (largest batch) · **Risk:** medium · split across 2-3 sub-PRs by sub-group.
+
+## Batch D — Shipped-content drift (commands / agents / workflows / templates)
+
+Stale cross-refs, renamed-command leftovers, contradictory instructions, nonexistent-file references.
+
+- **Commands/agents:** M1 (Copilot converter reads `allowed_tools:` no agent uses → no tool restrictions), M2 (`--copilot` missing from `--help`), M4 (upgrade never sweeps stale Claude skill shims), M7 (`--type plan|summary` vs code `plans`/`summaries`), M34 (`audit-deployment` false CRITICALs on healthy installs), M35/M36 (`experiment` stale status + `/pan:check` doesn't exist), M39 (`optimize` `trace status` → `trace current`), M40 (`retro` "modifies no files" false), M42 (`what-if` overwrite claim false), M43 (`help` phantom `--dry-run`), M44/M48 (`help`/help.md "42 commands" table omits 17 of 59 + embeds a count).
+- **Workflows/templates:** M45 (diagnose-issues wrong agent), M46 (exec-phase nonexistent continuation-prompt template), M47 (execute-plan resume detection dead), M49 (milestone-audit doubled `v{version}-v{version}` prefix), M50 (milestone-new cites nonexistent `/pan:discuss-milestone`), M51 (uppercase researcher filenames), M53 (resume-project case-mismatched substitution), M56 (verify-phase wrong glob).
+- **pan-zcode:** M80 (merge-approval instruction tells operator to use request id, not the nonce — no gated merge can be approved).
+
+**Effort:** M-L (high volume, mostly mechanical) · **Risk:** low.
+
+## Batch E — Docs correction batch
+
+Prose that contradicts code. Several overlap themes already fixed in Part 1 (balanced-tier, install-from-source, Codex skills path) — fold in the same corrections.
+
+- M63/M75 (ARCHITECTURE/USER-GUIDE stale map-codebase agent counts + UPPERCASE filenames), M64 (ARCHITECTURE "Codex hooks not supported" — contradicted), M65/M66/M67/M68 (CLI-REFERENCE phantom experiment/doc-lint flags), M69 (DEVELOPMENT local-testing refused by guard — same as H8), M70 (EXAMPLES milestone-done mid-milestone), M71/M76 (FAQ/README balanced=Opus — same as H5), M72 (FAQ Codex skills path — same as H9), M73 (MIGRATION impossible v2.10.0 rollback), M74 (TROUBLESHOOTING "no locking" — false since ADR-0030), M77/M78 (ADR-0027 status vs shipped; GLM/z.ai spec unmarked).
+
+**Effort:** M · **Risk:** low.
+
+## Batch F — Test-coverage holes
+
+- M57: global-install mode (`--global`/`--config-dir`/env-var) — no positive coverage on any runtime. Add a scenario test per resolution path.
+- M58: extract `pan-check-update.js`'s update-check logic out of the inline `node -e` string into a testable module, then cover it.
+- M59: `pan-context-monitor.js` (134 lines threshold/debounce/escalation) — add behavioral tests.
+
+**Effort:** M · **Risk:** low · high long-term value (these holes hid H1/H3-class bugs).
+
+## Batch G — Low tail (48) — hygiene sweep
+
+Embedded counts (count-SSoT violations), link rot, cosmetic drift. Largely mechanical; run the `doc-lint` count/flag scanners, delete embedded counts, fix stale links. Can piggyback on the next scheduled doc-audit rather than a dedicated PR.
+
+**Effort:** S-M · **Risk:** trivial.
+
+---
+
+## Suggested sequencing
+
+1. **A** (systemic parity lint — prevents recurrence, closes a class)
+2. **B** (dead gates — trust) + **C-security** (M22/M23/M79 injection/containment)
+3. **C-rest** (core correctness, 2-3 sub-PRs)
+4. **D** + **E** (drift + docs — high volume, mechanical, parallelizable)
+5. **F** (coverage holes)
+6. **G** (Low hygiene — fold into routine doc-audit)
+
+Each batch: fix → add/extend tests → `npm run test:all` green → live-verify installer-facing items in `d:\pantesting` → PR. Batches A–C are the real bug-fix value; D–G are drift/hygiene/coverage.