quant_engine/handoff-factor-diagnostics.md

# Quant OS factor diagnostics and strategy core increments

## 2026-10-04 seven-strategy core increment

The same owner, lifecycle, branch and PR #21 now include the independent seven-strategy calculation increment needed by platform Draft #102. The original 3c97102f CI-wait receipt has not been queried, resumed or treated as resolved. New code is being reviewed as a new candidate; previous fixed-archive platform consumers remain bound to their saved revisions.

The active source binding is the clean pinned49d0fc5929653a3a98a0edcf1da630237dcec770. Its core feature entry was reused; the delivery operation returned §16/§21 and those ranges were read. Resume/review parameters critical gpt-6-astra/xhigh passed; actual runtime remains unknown. Root remains sole code/Git writer. The business stage source is absent; no foreign stage ledger or task is borrowed.

Pure OHLC research now uses the existing daily ledger with a post-close policy and next-open execution. Seven signal definitions, explicit historical ATR stop, source/benchmark states, fee/cash timing, completed-lot FIFO accounting and a maximum100-combination real optimizer are implemented. Financial scope is L3; no source access, DB, migration, live execution, service reload or platform production admission is included. Precise methods, numeric guards and current evidence are in docs/strategy-research.md.

Actual tests first failed for missing APIs. Review and additional boundary tests reproduced and closed the documented unfilled-exit, numeric, fractional-holding and FIFO defects. Final focused run178 passed (93 new plus85 existing execution), then the entire core suite1186 passed in15.42s with1168 pandas deprecation warnings. Final nine-file Ruff and four-module mypy passed. Both reused reviewers passed their final relevant scopes; the final FIFO review independently reconciled fee-bearing add/partial-exit/full-exit PnL with final NAV and retained real tiny balances. The pure-library increment is code complete; clean-candidate save/delivery receipts follow at the final boundary. The broader platform/M0–M5 scope remains unfinished. Commit-budget continuation: this same delivery now includes an independently tested seven-strategy calculation increment plus necessary saved-revision receipts; preserve reviewed history rather than split or rewrite the delivery. Next: complete this new core candidate's applicable review and one central delivery boundary, then bind the saved source in the platform candidate task/artifact path without opening formal run/optimize.

The reviewed implementation was saved and pushed as `7d3e840483d7f3b5d7b2d987c95fb79a5dc4a63b` to the existing PR #21 (https://gitea.puyuanfh.cn/ageorge156/quant_engine/pulls/21), advancing the original 3c97102f candidate. This receipt-only checkpoint does not change the tested code. One explicit central candidate refresh and `ship --ready --confirm-l3` is the next boundary; no prior CI wait was polled or resumed. Push is a save checkpoint, and does not establish merge or source admission.

## Previous factor-diagnostics increment

Delivery: quant-os-factor-diagnostics-20261004; branch codex/quant-os-factor-diagnostics-20261004. Base is remote-confirmed origin/main 861c1e97a8bf1c3e962c4cd1ee88ef58e6b9ddd5. This is the independent core-repository increment consumed by the existing platform delivery/Draft #102, not a new platform branch or chat. Root in chat 01a0bc8f-dcaf-7452-9ab3-3215df6dfa97 is the sole core code/Git writer; reviewers are read-only. Primary, merged retrospective-v2 and old Alpha158 worktrees and their user files remain untouched.

The fixed framework source is 16e96351fbc5bd5a918c7deff556490bbc44fb99. Valid source cleanliness/version evidence was reused; actual core feature/worktree router ranges and applicable AGENTS were read. Resume/review model-policy critical gpt-6-astra/xhigh passed; actual runtime remains unknown. The core has no stage ledger at its declared path, so no foreign stage state is borrowed. Lifecycle created this task after policy-check; old completed trees were retained for ignored local data, no old lease/state was rewritten. The new .venv was created through lifecycle run with the existing frozen lock and dev extras, offline from local caches.

Scope is L3 mathematical behavior, pure memory/caller data. The new candidate module provides daily IC/RankIC summaries, keyed pooled correlation matrices and explicit-calendar forward-return labels. Production algorithm/source/PIT/execution qualifications remain unestablished and decision_eligible=false. Existing factor-library and governed factor-set contracts are unchanged. Reuse rationale and exact numerical/statistical assumptions are in docs/factor-diagnostics.md.

Evidence so far: initial47 tests failed for the missing API; first implementation46 passed and one strict floating-zero assertion failed, corrected to a stated1e-15 tolerance without clipping values. A genuine extreme-rank scaling regression was reproduced and fixed by ranking original values. An affine-equivalent IC case reproduced enormous IR/t from rounding dispersion, fixed with the public32eps resolution policy. Independent reviewer found two representable-offset Pearson errors; both were reproduced and fixed by translate-before-scale. Latest54 new plus16 old tests=70 passed; final whole-core run1093 passed in12.82s with1166 existing pandas deprecation warnings. Targeted Ruff and final module-only mypy passed. Reviewer rechecked the two numeric fixes with tiny independent samples, no new blocker.

The implementation was saved and pushed as 69383a30b079790018edb25679be08460a5ab17f in Draft PR #21 (https://gitea.puyuanfh.cn/ageorge156/quant_engine/pulls/21). The existing platform delivery now consumes that exact prepared archive through an unwired thin adapter. Its 25 tests call this real core; 9 isolated HTTP tests and 21 actual-component tests pass. The local browser displayed daily IC [1, -0.5, approximately 0], mean 1/6, actual daily pairs, pooled pairs, constant/missing reasons and holding-period valid days [3, 2, 1]; source failure cleared old results and explicit retry restored them. The prepared-report boundary rejects conflicting counts, intervals and method claims without recalculating statistics. Platform changes remain its separate, unmerged delivery; this is evidence for a consumer, not combined production acceptance.

The bounded pure-library increment is code complete and reviewed. The next operation is the central delivery gate for this core PR; its receipt, not this handoff, establishes Ready/CI/merge/main acceptance and cleanup facts. After successful core delivery, the platform will bind the accepted core source revision while retaining candidate and unestablished qualifications. Main full-suite evidence remains the unchanged 1093-pass candidate run above until the gate establishes its own validation receipt.

No source/provider/NAS query, production ETL, migration, release, deployment or trade occurred in this increment. Production data/source/PIT/algorithm admission is outside this pure-library delivery and remains explicitly unestablished. The broader platform/M0–M5 delivery remains unfinished. Reuse this same lifecycle/branch/PR; a push alone is not completion.