# Pass 3 — Reversibility classification

Date: 2026-05-16
Applies to: 14 EXECUTION_READY tasks only.

## Distribution

| Tier | Count | Tasks |
|---|---|---|
| HIGH | 4 | p0-freshness-dashboard, data-freshness-audit, scorecard-time-of-day-equity, pm-nightly-silent-failure-detection |
| MEDIUM | 8 | p0-orchestrator, p0-identity-fn, p0-param-surface-registry, refactor-scientific-overlay-js-to-load-from-system, strategy-vpi-rules, mgp-daily-stop-pipe, pm-nightly-task, scorecard-strategy-time-of-day |
| LOW | 2 | catalog-refactor-shared-json, add-user-role-approval |

**Healthy distribution** — 4 HIGH at the top means there's real autonomous-queue depth. The 2 LOW are both compounding/leverage-rich and correctly route to the supervised queue.

## Per-task classification

### HIGH

#### p0-freshness-dashboard
**Justification:** Adding/cronning a JSON file + adding a banner DOM block to `/system/index.html` is single-file rollback per concern. Consumers fall back to text "needs_db_tunnel" (already the current state). No schema migration; no widely-consumed contract change.
**Rollback:** Remove cron entry. Revert banner DOM injection on `/system/index.html`. Optionally delete the regenerated `/system/freshness.json` (it's a snapshot, not authoritative state).
**Mandatory verification checks:** N/A (HIGH).

#### data-freshness-audit
**Justification:** Read-only diagnostic. Produces a report at `/todos/audit/2026-05-16/freshness-audit.md`. No code changes; no schema changes.
**Rollback:** N/A — read-only.

#### scorecard-time-of-day-equity
**Justification:** Single chart axis change in `gen_scorecard.py`. Y values (cumulative P&L) stay identical; only X-axis label/order changes. Reference scorecard regen verifies.
**Rollback:** Revert the axis assignment in gen_scorecard.py. One-line change.

#### pm-nightly-silent-failure-detection
**Justification:** Adds health-check + alert step inside the nightly task SKILL.md. Existing nightly output unaffected by the additive check.
**Rollback:** Remove alert step from SKILL.md.

### MEDIUM

#### p0-orchestrator
**Justification:** New wrapper script (`/z/setup_daily_reporting/run_nightly.py`) installed alongside existing per-generator schedules. Cutover (disable individual triggers, enable orchestrator trigger) is the more sensitive part; before cutover, both run side-by-side. Per-property revert is local.
**Rollback:** Point Task Scheduler back to the per-generator entries; disable run_nightly.py trigger. Both sets of schedules are independent.

#### p0-identity-fn
**Justification:** Additive helper module (Python + PHP). Each consumer surface migrates independently — `gen_scorecard.py` first, then monthly, then accountability-v2, etc. Each surface revert is a local git revert.
**Rollback:** Revert generator commits one-by-one; helper module can remain unused.

#### p0-param-surface-registry
**Justification:** New index + per-strategy pages live in `/param-surface/` tree. Existing `/param-surface-db/` stays untouched throughout. Cutover to `/system/strategies.json` registry IDs is the LOW part — that's why the registry-only step is MEDIUM but the catalog rewire is its own LOW task (`catalog-refactor-shared-json`).
**Rollback:** Restore old `/param-surface/` index from `.bak`; keep VWAP grid as fallback.

#### refactor-scientific-overlay-js-to-load-from-system
**Justification:** JS module change in `_scientific_overlay.js`. Loaded fetch contract (the new dependency on `/system/strategies.json`) is the only new surface area. Per memory `project_canonical_taxonomy.md`, the JSON file already exists and matches the schema.
**Rollback:** Restore inline catalog in `_scientific_overlay.js` from git.

#### refactor-gen-scorecard-py-classify-entry-to-use-re (downstream of above)
**Justification:** Python function change in `gen_scorecard.py`. Old `classify_entry()` body trivially restored.
**Rollback:** Restore inline `classify_entry()` body from git.

#### split-td-vwap-from-vwap-pull-in-in-classifier (downstream)
**Justification:** Classifier change visible in scorecard pie; both batch + live evaluator affected. Two new tag values introduced.
**Rollback:** Collapse the two registry entries back; classifier returns to merged behavior.

#### strategy-vpi-rules
**Justification:** Edits `/system/strategies.json` to codify VWAP Pull-In gates/thresholds. `.bak` from 5/15 family migration provides one rollback step.
**Rollback:** Revert `/system/strategies.json` from `.pre-family-migration.bak` (preserved) or git.

#### mgp-daily-stop-pipe
**Justification:** Replaces hardcoded `$500` in MGP Max Risk calc with an API read from `/scorecards/api/user_settings.php`. Read fails → fall back to hardcoded value (defensive default).
**Rollback:** Restore hardcoded value in gen_mgp.py.

#### pm-nightly-task
**Justification:** PM SKILL.md update: bumps completion-status auto-update logic + adds visual progress section. Restoring prior SKILL.md is one file revert.
**Rollback:** Restore prior SKILL.md from git.

#### scorecard-strategy-time-of-day
**Justification:** Adds stacked bar + correlation matrix to scorecard. Existing pie unchanged.
**Rollback:** Remove new chart blocks from gen_scorecard.py.

### LOW

#### catalog-refactor-shared-json
**Justification:** Touches 5 surfaces (scorecards, strategy_library, live_classifier, param_surface, playbooks). The additive copy (`/system/strategies.json` already exists post-5/15 migration) is MEDIUM, but the **cutover** — making each of the 5 surfaces actually read from the canonical file — is LOW because:
1. If the canonical file has a subtly different schema than the inlined copy, all 5 surfaces misrender simultaneously.
2. Forward-deletion of the inlined copies eliminates the safety net.
3. Three surfaces (live_classifier, playbooks, param_surface) currently regenerate from their inlined copies — pre-existing tooling needs adaptation.

**Mandatory verification checks (per LOW rules):**
- `/system/strategies.json` schema is the canonical (post-5/15 migration; confirmed in `/system/` dir + memory `reference_canonical_taxonomy.md`).
- After each surface cuts over, visual comparison vs pre-change rendering shows no diff.
- Strategy IDs round-trip: scorecard pie label X = strategy_library card X = live_classifier tag X.
- No downstream regressions observed for 5 trading days after final consumer cuts over.

**Rollback:** Per-surface revert plan. Each surface can fall back to its prior inline catalog via `git revert <commit>`. The canonical JSON itself stays — it's already in use by /strategies/ rendering.

**Supervised: YES.** James pairs while each surface cuts over; verifies parity per surface.

#### add-user-role-approval
**Justification:** `ALTER TABLE main.users ADD COLUMN role`, `ADD COLUMN is_approved`. Reversible (drop columns) but once consumers (gen_scorecard.py, accountability-v2, strategy-upload-ui) start reading the columns, downgrade requires removing the read paths too. Touches scorecards + accountability_hud + strategy_library (the 3 downstream consumers documented in `unlocks`).

**Mandatory verification checks:**
- After migration, `SELECT role, is_approved FROM main.users WHERE id=2` returns `('admin', true)` for James.
- gen_scorecard.py + accountability-v2 + strategy-upload-ui can read role from a single join.
- A test trader row with role='trader', is_approved=false renders correctly on each consumer.
- No downstream regressions observed for 5 trading days after deployment.

**Rollback:** `ALTER TABLE main.users DROP COLUMN role, DROP COLUMN is_approved`. Any data set on those columns is lost; document the assumption that they're not load-bearing data yet (just metadata).

**Supervised: YES.** Schema migration with downstream consumers. Belongs in the paired queue.

## Disqualified-from-EXECUTION_READY due to verification gap

Per the Pass 3 rules: "If a LOW_REVERSIBILITY task can't have these added [verification + rollback], it doesn't qualify for EXECUTION_READY — it's downgraded to STILL_PENDING_REAL with a note that verification criteria need definition before autonomous execution."

After adding the mandatory checks above, **all LOW tasks retained EXECUTION_READY status**. No disqualifications.

## Notes on the distribution

- The 4 HIGH tasks are all safe to attempt in any order on a Saturday. They're the bottom of the Tier 4 work in the weekend pick list, but they're also where the autonomous queue gets its volume.
- The 8 MEDIUM tasks include the 3 P0 compounding-infra wins (p0-orchestrator, p0-identity-fn, p0-param-surface-registry). MEDIUM is the right call for each — additive helper modules with iterative consumer cutover.
- The 2 LOW tasks both meet verification criteria; both belong in the supervised queue.
- **No EXECUTION_READY task currently fails the LOW-task verification requirement.** The patch.json adds the verification checks as `done_when` entries.
