[initial-d]

Summary

Both public validation scripts used the forward-label mask (t and t+1 observed) for portfolio selection and out-of-sample prediction. A stock halted tomorrow therefore disappeared from today's portfolio. This separates signal-date eligibility from historical label eligibility in both paths; training and trailing IC retain forward-label filtering, while selection and inference use today's mask.

A same-input synthetic replay shows the affected equal-weight position moving from zero to 1/12 on the day before the halt, with zero on the actual halt. Returns change in both directions; this is a causality fix, not performance evidence. New metadata records availability_protocol, and archived reports and README headline results link the protocol caveat.

Evidence and limitations

• docs/audits/public-revalidation-20261017/ preserves the predeclared scope, environment and dependency freeze, exact commands, checksums, old-code assertion failures, and replay script/results.
• The intended public A-share rerun was data-blocked: AkShare single-attempt probes for 600519 and 000001 both disconnected without data. The log and timestamps are preserved. No public-data result or proprietary-paper reproduction is claimed.
• No raw market data is committed. Execution across halts/limits remains simplified; equal evaluation windows and real all-factor/MLP coverage need separate validation.

Validation

• Both portfolio causality regressions fail on base ccdcf16 for the affected zero holding and pass after the fix.
• Full suite: 143 passed.
• Ruff on src/tests and the affected scripts: passed.
• Synthetic CPU CLI demo: completed.
• Same-input replay and generated synthetic report audit: passed.
• Documentation links, evidence JSON, checksums, and git diff --check: passed.

Type: bug fix, tests, documentation. Solo contribution; no external messages or market data publication.