create-tradejs 3.1.22 → 3.1.23

package/dist/skill-bundle/.codex/skills/strategy-improvement-research/SKILL.md

@@ -41,6 +41,18 @@ core/gate/tests, and commit the selected source candidate locally. Do not push,
 publish packages, edit the Project’s production composition, deploy, start a
 forward test, or change live risk. Those belong to `$strategy-forward-start`.
 
+## Orchestration boundary
+
+This skill owns the complete improvement lineage: prior-candidate inventory,
+hypothesis-family choice, trial budget, parent/child decisions, candidate
+selection, and final handoff. Delegate each preregistered core implementation
+and backtest to `$strategy-backtest-research`; that skill returns reconciled
+experiment evidence and does not choose the next candidate. After freezing the
+core finalist/export, delegate deterministic-gate analysis to
+`$ai-train-local-research`; it must not reopen core selection. Read those
+specialist skills when their stage begins instead of duplicating their command,
+metric, or reporting contracts here.
+
 ## Required contour
 
 1. Start a new immutable research lineage. Freeze data bounds, point-in-time
@@ -56,6 +68,8 @@ forward test, or change live risk. Those belong to `$strategy-forward-start`.
    with evidence-driven children and use remaining slots for direction-policy
    or Pareto rescue. The default cap is 12 genuinely new behaviors. An audit,
    parser fix, no-op, or rescoring of an old behavior does not consume a slot.
+   Execute each frozen experiment through `$strategy-backtest-research` and
+   consume its verified result as evidence for the next orchestration decision.
 5. Do not stop at the audit, baseline, first failed round, or a sparse recent
    tail. Stop only when a reproducible best candidate is frozen, the fresh
    budget is exhausted, or every remaining family has a recorded hard causal
@@ -63,7 +77,10 @@ forward test, or change live risk. Those belong to `$strategy-forward-start`.
 6. Keep one chronological tail sealed during discovery when coverage permits.
    Open it once for the final selected behavior. Track all exposed tests for
    multiple-testing/deflated-Sharpe interpretation.
-7. Run package formatting, typecheck, tests, and build in the lineage worktree.
+7. Freeze the selected core/export before opening deterministic-gate research.
+   Run the gate stage through `$ai-train-local-research`; do not retune or
+   relabel the raw-core result inside gate tooling.
+8. Run package formatting, typecheck, tests, and build in the lineage worktree.
    Commit only the selected candidate and its tests on that worktree branch;
    preserve rejected experiments as immutable evidence, not source clutter.
 

package/dist/skill-bundle/.codex/skills/strategy-release/SKILL.md

@@ -0,0 +1,22 @@
+---
+name: strategy-release
+description: Deprecated compatibility entrypoint for the former monolithic TradeJS strategy workflow. Use a focused strategy-candidate, strategy-improvement, strategy-period, strategy-forward, or strategy-risk skill instead.
+---
+
+# Strategy Release (Deprecated)
+
+Do not run the former all-in-one research and production contour. Route the
+request to exactly one focused skill:
+
+- latest selected metrics: `$strategy-candidate-report`;
+- candidate versus production: `$strategy-candidate-compare`;
+- improvement analysis only: `$strategy-improvement-plan`;
+- new core + deterministic-gate research: `$strategy-improvement-research`;
+- frozen candidates on a new period: `$strategy-period-revalidate`;
+- start or replace a risk-1 forward test: `$strategy-forward-start`;
+- inspect a running forward test: `$strategy-forward-status`;
+- increase only `MAX_LOSS_VALUE`: `$strategy-risk-scale`.
+
+Never combine research, candidate selection, publication, deployment, and risk
+scaling merely because this legacy name was invoked. Ask for the intended
+focused action only when it cannot be inferred from the user’s wording.

package/dist/skill-bundle/.codex/skills/strategy-release/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Strategy Release (Deprecated)"
+  short_description: "Deprecated compatibility router"
+  default_prompt: "Use $strategy-release only as a compatibility router to the single focused TradeJS strategy skill that matches the requested action."

package/dist/skill-bundle/.codex/skills/strategy-release/references/diagnose-live.md

@@ -0,0 +1,126 @@
+# Diagnose live behavior
+
+Use this mode to classify observed live behavior of an already released core
+plus deterministic gate composition. Do not tune or release a replacement in
+the same lineage.
+
+## 1. Bind the exact released composition
+
+Load the immutable release record and verify:
+
+- core config snapshot and fingerprint;
+- deterministic gate fingerprint and `MIN_AI_QUALITY`;
+- runtime mode, strategy code/git SHA, connector, interval, and ticker scope;
+- fees, slippage, entry delay, risk settings, and context-provider settings;
+- release acceptance bounds, terminal windows, and forward-incubation cutoff.
+
+Require the runtime artifact to embed deployment snapshot schema v2. Every
+scoped evaluation, signal, trade, and persisted scope must carry runtime lineage
+schema v3 matching the embedded deployment/account, strategy revision,
+strategy package version, complete dependency versions, and runtime package
+version. Track `MAX_LOSS_VALUE` separately: different values are allowed,
+produce immutable `L` evidence, and require all monetary observations to be
+normalized by `runtime / release` risk scale. Missing/invalid risk scale yields
+`INSUFFICIENT_EVIDENCE`; missing, conflicting, or different current lineage is
+`RUNTIME_DIVERGENCE`.
+
+Do not accept or translate another runtime evidence schema. Do not reconstruct
+composition from `strategyConfigs`, mutable Redis state, research fingerprints,
+or current source checkout metadata.
+
+Do not infer live execution from the current local Redis config. Production
+signals may run on another server. Request or inspect remote runtime artifacts
+when local evidence is not the source of truth.
+
+## 2. Freeze the incident window and evidence
+
+Freeze a half-open incident window `[start, end)`, affected symbols, runtime
+trade/signal/evaluation ids, closed-candle timestamps, gate decisions/reasons,
+order attempts/fills, fees/funding/slippage, exits, daemon lifecycle, and data
+freshness. Hash every input before analysis.
+
+Use only closed candles and signal-time-causal context. Preserve missing skip
+evidence as missing; do not treat absent detailed skip records as proof that an
+evaluation did not run.
+
+## 3. Establish parity before performance diagnosis
+
+Compare runtime with replay/backtest for the same composition and incident
+window:
+
+- effective config and fingerprints;
+- candle boundary and restored state/checkpoint;
+- candidate signal and deterministic gate decision;
+- allocator/risk/order status;
+- entry/exit timestamps, prices, reason, fees, funding, and slippage.
+
+Any historical backtest used for context must include `--cacheOnly` and the
+maximum common cached window shared by the comparison. Do not update history to
+make replay match live.
+
+Classify unmatched items by evidence, not by assumption: no runtime evaluation,
+gate/policy block, completed signal without fill match, runtime-only candidate,
+backtest-only candidate, price/timestamp mismatch, state/config mismatch, or
+context-data mismatch.
+
+## 4. Separate divergence from performance
+
+If a material runtime/replay/config/context mismatch explains the incident,
+return `RUNTIME_DIVERGENCE`. Do not judge core generalization from a non-parity
+sample.
+
+If parity holds, compare the incident and forward-incubation observations with
+the frozen release distribution:
+
+- `ALL`, `LONG`, and `SHORT` N/PnL/PnL-trade/PF/WR/DD/cadence;
+- matched release regimes and causal setup cohorts;
+- symbol/event concentration and simultaneous batches;
+- expected loss streak, losing-period, and drawdown bounds;
+- data lag and independent timestamp/event support.
+
+Return `EXPECTED_DRAWDOWN` only when parity is established and the observation
+remains within preregistered bounds. Return `GENERALIZATION_FAILURE` only when
+parity is established and adequate new post-cutoff evidence breaches the frozen
+generalization rule. Otherwise return `INSUFFICIENT_EVIDENCE`.
+
+## 5. Keep comparisons advisory
+
+Do not alter core or gate thresholds in diagnose mode. Record a proposed causal
+hypothesis as a new future release lineage.
+
+If `llmComparison=ai-approved`, evaluate only deterministic-gate-approved live
+or replay rows with the explicitly configured provider/model. Store agreement,
+contradiction, and reasons as advisory evidence. Do not use the LLM to relabel
+runtime divergence, override deterministic decisions, or trigger orders.
+
+## Suggested commands
+
+Choose commands supported by the captured evidence and record every flag:
+
+```bash
+yarn replay --startTime <start-ms> --endTime <end-ms> --cacheOnly
+
+yarn runtime-parity --startTime <start-ms> --endTime <end-ms> --details
+
+yarn ai-train --strategy <Strategy> --file <released-export-part1.jsonl> \
+  --localOnly --json -n 0 --terminalWindows=90,30,7
+
+yarn runtime:scorecard \
+  --strategy <Strategy> \
+  --runtimeEvidence <verified-runtime-evidence.json> \
+  --replayEvidence <replay-runtime-evidence.json> \
+  --calibration <execution-calibration.json> \
+  --prospectiveEvidence <raw-core-gate-regime-summary.json> \
+  --releaseManifest <release-envelope.json> --diagnosisDays <7|30|90> \
+  --strategyReleaseRoot data/strategy-release
+```
+
+Do not run a broad historical backtest unless it is needed to compare the
+incident with the frozen release bounds; when needed, use the exact released
+config and maximum common cached window only.
+
+## Diagnostic ready prompt
+
+```text
+Use $strategy-forward-status for <Strategy>. Bind release record <path-or-id> to runtime evidence <path-or-id>, freeze incident window <start>..<end>, and check closed-candle/config/state/context/gate/execution parity before performance. Use only --cacheOnly historical comparisons over the maximum common cached window. Set llmComparison=<off|ai-approved>. Return exactly one evidence-status verdict and make no runtime or trading changes.
+```

package/dist/skill-bundle/.codex/skills/strategy-release/references/direction-policy.md

@@ -0,0 +1,141 @@
+# Direction-policy checkpoint
+
+Use this checkpoint after the core rescue board and before concluding that no
+composition can be salvaged. It tests whether a useful raw direction is being
+obscured by the opposite direction or by the current deterministic gate.
+
+## Invariants
+
+- Keep both directions enabled in the authoritative raw-core run and retain its
+  ALL/LONG/SHORT evidence.
+- Never rewrite a failed raw-core result as profitable by deleting a side from
+  its metrics.
+- Treat direction policy as explicit composition logic. Freeze, hash, and test
+  `both`, `long_only`, and `short_only` policies like other deterministic-gate
+  variants.
+- Prefer a deterministic-gate direction block to changing the core side toggle:
+  it preserves raw counterfactual telemetry and makes the suppression visible.
+- A core side toggle may be used only as an explicit equivalence candidate. It
+  must reproduce the corresponding gate-blocked entries and be separately
+  authorized before runtime configuration changes.
+- Direction containment may produce a composition finalist even when the raw
+  ALL cohort failed. It does not waive maximum-covered economics, cost,
+  support, concentration, drawdown, or causality rules for the retained side.
+  Terminal calendar rows remain historical-readiness diagnostics.
+
+## Required trigger table
+
+Evaluate every row before returning `UNSUITABLE_FOR_CURRENT_MARKET` or
+`STOP_RESEARCH`:
+
+Create the preregistered input and run the deterministic classifier before
+freezing variants:
+
+```bash
+node .codex/skills/strategy-release/scripts/direction-policy-checkpoint.mjs \
+  --input <direction-policy-input.json> > <direction-policy-decision.json>
+```
+
+The input contains raw ALL/LONG/SHORT metrics, the frozen useful-side rule, and
+current gate-approved side counts when available. Hash the input and output.
+Do not override a required checkpoint in prose.
+
+| Raw/gate evidence                                                                                                       | Required action                                                                                                                                            |
+| ----------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| One raw side has positive PnL and PF above 1 with adequate support, while the other side is the dominant aggregate loss | Run the five-variant direction-containment gate checkpoint. A no-gate-stop verdict is invalid.                                                             |
+| A useful raw side receives zero or negligible current-gate approvals                                                    | Run side pass-through and direction-aware rescue variants.                                                                                                 |
+| The retained side is profitable long-window but fails a selection-grade recent guardrail or cost stress                 | Still evaluate the frozen direction policy; classify it as diagnostic or prospective-only unless it passes the release rule. Do not call it runtime-ready. |
+| Both raw sides fail the preregistered side edge rule                                                                    | No direction salvage is required, but the full diagnostic report remains mandatory.                                                                        |
+| Side support is too sparse for release inference                                                                        | Keep the policy research-only or micro-forward eligible according to the verdict contract; do not manufacture historical support.                          |
+
+`Useful raw side` is frozen before the checkpoint. By default it requires
+positive full-development PnL, PF greater than 1, positive PnL/trade, and the
+preregistered minimum support/cadence. A study may use a stricter rule but may
+not invent a looser rule after seeing the result.
+
+## Five frozen variants
+
+Freeze these variants before opening tuning/test economics:
+
+1. current deterministic gate with both directions;
+2. hard block of the failing direction (`long_only` or `short_only`);
+3. retained-side raw pass-through plus the same failing-direction hard block;
+4. current gate plus one rounded causal repair pocket for the failing direction;
+5. direction-aware replacement using the best preregistered policy per side,
+   including a hard block where that is the frozen candidate.
+
+When the problem is instead a profitable side hidden by the gate, variants 2
+and 3 become target-side raw pass-through and pass-through plus a protective
+exclusion, matching the release workflow's recovery semantics. Record which
+case is active so the same label cannot change meaning after outcomes are known.
+
+For each variant report:
+
+- raw and approved ALL/LONG/SHORT metrics;
+- approval share and reject reasons by side;
+- full, 3y, 4y, maximum-covered, 365d, 180d, 90d, 30d, and 7d windows;
+- cost stress, folds/months, regimes, concentration, event fan-out, loss streak,
+  and drawdown;
+- retained-side support and non-target removal/occupancy effects;
+- train, tuning, and untouched-test partitions.
+
+If a blocked side has zero approved rows, report explicit zero rows rather than
+omitting it. The raw side remains visible beside the approved side.
+
+## Decision examples
+
+### Positive LONG, losing SHORT
+
+Raw LONG `PnL > 0`, `PF > 1`; raw SHORT supplies most aggregate loss. Test
+`long_only` and the remaining four variants. If LONG then fails a
+selection-grade terminal cohort or cost stress, reject runtime readiness for
+those reasons. A terminal loss alone may still leave the immutable policy
+eligible for a risk-1 prospective test; a cost or portfolio-risk failure does
+not. Keep underpowered/diagnostic 90d/30d/7d rows visible as `n/a` or diagnostic
+evidence; do not let them reject the policy. Do not claim that SHORT made
+testing impossible.
+
+### Profitable SHORT hidden by the gate
+
+Raw SHORT passes its edge rule but current qN+ approves no SHORT rows. Test
+SHORT pass-through, a causal SHORT pocket, protected SHORT, and the
+direction-aware replacement. A zero-approval current gate is not a final
+market-unsuitable result.
+
+### Both sides weak
+
+If neither side passes the frozen useful-side rule, record `no_side_salvage`.
+Still generate the complete raw-core and current-gate diagnostic reports and
+chart before the verdict.
+
+## Required artifact
+
+Persist a machine-readable checkpoint:
+
+```json
+{
+  "schema": "tradejs-direction-policy-checkpoint/v1",
+  "trigger": "losing_side_contamination|profitable_side_hidden|no_side_salvage",
+  "rawCoreResearchId": "<id>",
+  "usefulSideRule": {},
+  "rawSideStatuses": {
+    "LONG": "useful|failed|insufficient",
+    "SHORT": "useful|failed|insufficient"
+  },
+  "variants": [
+    {
+      "id": "<id>",
+      "policy": "both|long_only|short_only|direction_aware",
+      "configSha256": "<sha256>",
+      "resultSha256": "<sha256-or-null>",
+      "status": "complete|invalid|not-run",
+      "decision": "retain|reject|diagnostic"
+    }
+  ],
+  "selectedPolicy": "both|long_only|short_only|direction_aware|null",
+  "reason": "<bounded evidence statement>"
+}
+```
+
+Hash and cite it from the final release note and decision input. A required but
+missing checkpoint makes the release evidence incomplete.

package/dist/skill-bundle/.codex/skills/strategy-release/references/directional-parameter-split.md

@@ -0,0 +1,93 @@
+# Directional parameter split
+
+Use this checkpoint when one isolated global parameter change helps one
+direction and harms the other. It converts a measured side conflict into a
+causal core experiment; it is not permission to duplicate every config field.
+
+## Trigger
+
+Run the classifier after a complete reconciled matched comparison:
+
+```bash
+node .codex/skills/strategy-release/scripts/directional-parameter-checkpoint.mjs \
+  --input <directional-parameter-input.json> \
+  > <directional-parameter-decision.json>
+```
+
+Classify each side with the preregistered side rule as `improved`, `worsened`,
+`neutral`, or `inconclusive`. Require adequate independent support, the intended
+trace transition, and a one-parameter config delta. A split is required only
+for `LONG improved / SHORT worsened` or the mirrored conflict.
+
+Do not trigger from aggregate PnL alone, one profitable symbol/event, an
+outcome-derived cohort, or an unexplained occupancy change. When the original
+candidate changed several fields, run an ablation first.
+
+## Implementation contract
+
+Keep the original field as the backward-compatible fallback and add optional
+directional overrides:
+
+```text
+PARAM
+PARAM_LONG
+PARAM_SHORT
+```
+
+Resolve only `undefined` to the fallback. Values such as `0`, `false`, and an
+empty string are valid explicit overrides. With both overrides absent, old
+configs, decisions, payloads, and state keys must remain exact.
+
+Choose the implementation according to where the value acts:
+
+- `decision_time`: resolve after direction is known; use a normal fallback
+  override.
+- `detector_state`: a shared detector cannot simply switch values after a
+  signal appears. Use isolated replay-safe LONG and SHORT detector state or
+  reject the split as architecturally unsafe.
+- `shared_lifecycle`: add the override, but preregister occupancy/cooldown and
+  opposite-side identity guardrails because one side can change the other's
+  opportunity set.
+
+Include the effective LONG and SHORT values in detector/execution config
+identity wherever they can alter replay state or decisions. Add focused tests
+for implicit fallback parity, explicit LONG/SHORT resolution, zero/false
+overrides, config isolation, replay reconstruction, and same-timestamp behavior.
+
+## Minimal research ablation
+
+The global candidate is already evidence. Spend existing child/rescue slots,
+not extra trials, on the smallest discriminating set:
+
+1. exact legacy control with no overrides;
+2. target-side-only override, keeping the opposite side at control;
+3. combined best-per-side override only when both side values have independent
+   prior evidence.
+
+Use an explicit equal-overrides parity case in unit tests; backtest it only when
+runner/config resolution parity is uncertain. Do not automatically test a full
+Cartesian grid.
+
+For the target side require the frozen improvement rule. For the non-target
+side require exact signal/trade identity when architecture guarantees
+independence. Otherwise report matched/added/removed identities, occupancy
+spillover, N/cadence, PnL/trade, PF, WR, and side-only DD under a preregistered
+non-regression rule. Keep aggregate portfolio DD as a separate guardrail.
+
+## Interpretation
+
+- A successful target-only override supports a genuine directional parameter
+  asymmetry and may be carried forward as core behavior.
+- If the target side improves but the opposite side changes through occupancy,
+  judge the whole causal composition rather than claiming isolation.
+- If equal explicit overrides do not reproduce the global candidate, fix config
+  resolution/state identity before using any economics.
+- If a detector-state split needs duplicated engines whose state cannot be
+  replayed exactly, retire the split rather than accepting runtime divergence.
+- Preserve the original global parameter even after promotion so existing
+  configs remain reproducible and future configs can still choose symmetric
+  behavior.
+
+Persist the classifier input/output, parameter semantics, fallback contract,
+resolved configs, implementation tests, side identities, and decision in the
+immutable family handoff.

package/dist/skill-bundle/.codex/skills/strategy-release/references/evidence-limitations.md

@@ -0,0 +1,76 @@
+# Evidence limitations and action ceiling
+
+Classify an evidence problem before deciding whether to stop work. Do not use
+the generic word `contamination` for limitations with different consequences.
+
+## Hard invalidation
+
+These conditions invalidate selection economics and stop the affected run:
+
+- outcome, future, or delayed-execution fields entered a signal-time decision;
+- export and Redis N/W/L do not reconcile;
+- the manifest is partial, failed, OOM, or belongs to another run;
+- candidate and control use different causal windows, costs, or data revisions;
+- a strategy state/reset leak mixes candidate cells;
+- required candles or causal context are absent for the evaluated timestamps.
+
+Repair or rerun the invalid evidence. A hard-invalid run cannot rank a
+candidate, but the failure does not excuse skipping other valid families.
+
+## Research-grade provenance limitation
+
+These conditions reduce the claim ceiling but do not stop paired research:
+
+- the current deployable symbol cohort is applied retrospectively because
+  effective-dated exchange membership is unavailable;
+- a chronological tail was exposed before variant freeze;
+- an old uncommitted patch is not byte-reconstructable, while its causal
+  mechanism can be implemented again as a new candidate;
+- remote deployment/account bindings are absent on the local research machine;
+- prospective independent-event support is not yet available.
+
+For a retrospective current-universe cohort:
+
+1. Label the estimand honestly: "performance of today's deployable cohort over
+   its available cached history", not historical exchange-wide performance.
+2. Keep control and candidates on the identical cohort/window and emphasize
+   matched setup/trade deltas over absolute PnL.
+3. Run membership sensitivity when the cache permits it: stable incumbent
+   symbols, minimum cached-history-age cohorts, and leave-one-symbol/event
+   concentration checks. Never invent listing/delisting dates.
+4. Use the result to select a prospective candidate. It cannot by itself
+   establish unconditional historical robustness or `READY_FOR_RUNTIME`.
+5. If the bounded composition otherwise qualifies, prepare or start the
+   authorized risk-1 micro-forward. Prospective current-universe evidence is the
+   correct way to resolve this limitation; waiting for a perfect historical
+   membership archive is not.
+
+An exposed holdout follows the same action rule: it cannot establish a sealed
+historical release claim, but it may select one immutable prospective
+composition. Do not tune again after that selection.
+
+## Operational limitation
+
+Missing local runtime bindings, credentials, or server Redis records do not
+block local research. Produce the portable handoff and mark it
+`requiresRuntimeBinding=true`. Resolve the binding only on the runtime server.
+
+## Action matrix
+
+| Limitation                         | Continue core/gate research           | Historical READY claim | Risk-1 micro-forward                    |
+| ---------------------------------- | ------------------------------------- | ---------------------- | --------------------------------------- |
+| Causal leakage/lookahead           | No, repair evidence                   | No                     | No                                      |
+| Reconciliation/partial run         | No, rerun                             | No                     | No                                      |
+| Retrospective current universe     | Yes, paired/sensitivity               | No                     | Yes                                     |
+| Exposed holdout                    | Yes, without retuning after selection | No                     | Yes                                     |
+| Missing old patch, mechanism known | Yes, new implementation lineage       | No claim from old run  | Yes after new evidence                  |
+| Missing local server binding       | Yes                                   | Unaffected             | Portable handoff locally; bind remotely |
+
+Record one of these ceilings in the release progress artifact:
+
+- `historical_ready_eligible`;
+- `micro_forward_only`;
+- `invalid_evidence`.
+
+The ceiling controls claims, not whether Codex performs the next bounded
+research action.

package/dist/skill-bundle/.codex/skills/strategy-release/references/evidence-retention.md

@@ -0,0 +1,157 @@
+# Evidence retention
+
+Preserve enough immutable evidence to reproduce every displayed metric and
+understand every trial after Redis, caches, and exported files are deleted.
+
+## Immutable lineage
+
+Create a new release or diagnostic record for each question and immutable
+lineage. Never append a new run, changed composition, opened holdout, or revised
+decision to an older result record.
+
+Store internal research under the existing repository note contract:
+
+```text
+$PROJECT_CWD/notes/<Strategy>/YYYY-MM-DD-<release-or-diagnose-slug>.md
+```
+
+Keep `notes/` ignored and never stage or force-add it. Embed the complete
+secret-free evidence in the note; mutable Redis keys and artifact paths are
+inventory only.
+
+## Required release evidence
+
+Retain:
+
+- every `tradejs-release-progress/v2` input/output pair, including objective,
+  historical-candidate revalidation, artifact-backed round/rescue evidence, and
+  selected-composition lineage, evidence ceiling, and the next stage that
+  prevented a premature final verdict;
+- the evidence-limitation classification (`hard invalidation`,
+  `research-grade provenance`, or `operational`) and sensitivity results;
+- the trading thesis, opportunity map, hypothesis-role portfolio, and per-round
+  belief-ledger updates that explain why each experiment was worth running;
+- experiment id, question, mode, preregistered acceptance rules, verdict, and
+  timestamp when each result partition was opened;
+- the hashed strategy-history hypothesis inventory: behavior-relevant commits
+  and dirty patches, semantic/config hashes, evidence matches, classification,
+  bridge status, selected anchors, excluded duplicates/refactors, and unresolved
+  backlog;
+- the prior-result bridge table with source/config/window/universe/cost lineage,
+  ALL/LONG/SHORT metrics, comparability class, rerun ids, and explanation for
+  every material regression or reproduction;
+- append-only trial ledger containing the common control, all 3 causal families,
+  three refinement-round indices, rescue-board index, parent research IDs, and
+  every attempted candidate up to the 18-variant cap, with status, resolved
+  config, selection rank, cadence region, mechanism verdict, and rejection
+  reason;
+- per-round authoritative metrics, setup matching, matched/added/removed/
+  changed-outcome and occupancy cohorts, trace-funnel/skip deltas, causal
+  regimes, cost stress, statistical guardrails, predicted-versus-observed
+  effects, and the evidence rationale for each next-round child variant;
+- each round's hashed causal-handoff payload, including direct parents,
+  eligible carried control, mechanism verdict, failure mode, family decision,
+  and preregistered primary/falsification or refinement/robustness child deltas;
+- the post-round-3 rescue board: complete candidate frontier, cadence-diverse
+  seed-selection inputs, each seed's dominant failure, one predicted rescue
+  delta per seed, rescue result, and the hard reason for any unused slot;
+- the mandatory direction-policy checkpoint: frozen useful-side rule, raw side
+  statuses, trigger, five gate variants, raw-versus-approved ALL/LONG/SHORT,
+  policy/config/result hashes, and selected/rejected reason;
+- every triggered directional-parameter checkpoint: one-field attribution,
+  side effects/support, resolution mode, classifier hashes, legacy fallback,
+  target-only/combined ablations, config/state isolation tests, and non-target
+  identity or occupancy-spillover evidence;
+- selected isolated-long finalist and final core-plus-gate composition;
+- exact git SHA/dirty diff inventory, config/gate/context fingerprints, and
+  tool/metric implementation SHA;
+- ordered ticker universe, checksum, eligible/raw counts, connector, interval,
+  exact `[start,end)`, maximum-common-cache coverage proof, and point-in-time
+  membership lineage;
+- fees, slippage, entry delay, `MAX_LOSS_VALUE`, AI mode, quality threshold,
+  provider/context settings, and BOTH-direction proof;
+- exact commands, run ids, manifest status, planned/completed/error/OOM counts,
+  config ids, export merge/part ids, hashes, and Redis reconciliation;
+- complete machine-readable ALL/LONG/SHORT metrics for full and required
+  terminal/cold-start windows, including zero-trade cohorts;
+- the full diagnostic matrix even when no finalist qualifies: authoritative
+  control, best aggregate, best LONG, best SHORT, and every rescue/policy
+  attempt over full/3y/4y/max-covered/365d/180d/90d/30d/7d windows;
+- control/candidate matched, removed, added, changed-outcome, occupancy, regime,
+  month, symbol, event, concentration, and capacity evidence;
+- gate train/tuning/test boundaries, feature provenance, threshold rounding,
+  sealed-test flag/open timestamp, ablation, support, and one-round selection
+  decision; record any accidental early test exposure permanently;
+- the complete `$ai-train-local-research` report sections and structured
+  full-period `ai-train --localOnly --chart -n 0` output, labelled
+  `diagnostic-only` when no composition qualifies;
+- LLM comparison scope/provider/model/prompt/cost/fingerprint when enabled,
+  explicitly labelled advisory.
+
+Preserve risk-scale changes as compact permanent `L` markers. Do not fork or
+hide the unchanged core + gate evidence timeline merely because
+`MAX_LOSS_VALUE` changed; retain the old/new values, timestamp, source artifact,
+and normalization ratio.
+
+Do not overwrite rejected hypotheses. Preserve partial/OOM/error runs with an
+invalid-for-selection label so they are not silently retried as new evidence.
+
+## Required diagnose-live evidence
+
+Retain:
+
+- referenced immutable release id and exact composition fingerprints;
+- incident `[start,end)`, remote/local source authority, collection timestamp,
+  affected symbols, runtime ids, signal/evaluation/analysis/order/trade keys;
+- closed candles, detector state/checkpoint identity, baseContext/gate inputs,
+  gate decision and reasons, allocator/risk/order statuses;
+- requested and actual entry/exit timestamps/prices, exit reason, quantity,
+  fees, funding, spread, impact, delay, slippage, and realized PnL;
+- replay/backtest command and cached-coverage proof, match tolerance, matched and
+  unmatched classifications, nearest candidates, and per-field deltas;
+- forward-incubation cutoff, independent event count, ALL/LONG/SHORT and regime
+  comparison with frozen release bounds;
+- verdict precedence applied and unresolved evidence gaps.
+
+## Retention status
+
+Use one of:
+
+- `complete`: the record embeds all configuration, structured metrics, lineage,
+  and verdict evidence.
+- `partial`: diagnostic material is useful but cannot reproduce every claim.
+- `blocked`: a named missing or invalid input prevents the requested verdict.
+- `legacy-partial`: historical evidence predates the contract and must not be
+  filled from current defaults.
+
+Never mark evidence complete merely because a command exited zero. Verify the
+finished manifest, hashes, counts, reconciliation, and machine-readable metrics.
+
+## Forward incubation
+
+Write new post-cutoff observations to a new immutable record. Do not reopen the
+selection/test tail or mutate the release note. Advisory/shadow candidates may
+log counterfactual decisions and LLM comparisons, but they must not change
+orders, runtime config, risk, daemon state, or promotion status without explicit
+approval and a separate release decision.
+
+## Storage tiers and cleanup
+
+Use the default tiering unless the user later changes it:
+
+- operational Redis evidence: 3 days;
+- reproducible verbose payloads: 14 days;
+- verified, aggregated runtime bundles: 90 days;
+- compact trial ledgers, release manifests, outcomes, gate disagreements,
+  diagnoses, and chart markers: permanent.
+
+Review the plan before applying it:
+
+```bash
+yarn strategy:release retention --input <retention-inventory.json>
+yarn strategy:release retention --input <retention-inventory.json> --apply
+```
+
+The first command is a dry run. The planner must keep unverified or unaggregated
+evidence regardless of age so cleanup cannot destroy the only unresolved source
+artifact.