fix(statements): correct + unpublish the 2026-07-28 weekly (unresolved window, not a 0% hit rate)

[tkowalczyk]

2026-07-28-weekly.md is live at draft: false claiming:

ogsfrompoly weekly statement — 263 alerts, hit rate 0.00.

To a reader that says 263 calls, all wrong. In fact none of the 263 had
resolved. This PR corrects the numbers and takes the claim off the live site.

What changed

field	before	after
summary	… 263 alerts, hit rate 0.00.	… 263 alerts, outcomes pending (0 of 263 resolved).
hypothetical_pnl_usd	0.0	199.11
top_wallets	[]	3 entries
draft	false	true — see below
alert_count · period_* · hit_rate · title	—	unchanged

alert_count: 263 reproduces exactly against the same
2026-07-22 → 2026-07-28 window, which is what confirms the window is identical
and only the reporting of it was wrong.

Why it lands as a draft rather than a corrected publication

CI caught what the correction could not fix, and it was right to. With
hypothetical_pnl_usd restored to its true value, findPublishedInconsistencies
fires:

alert_count=263 with hit_rate=0 but hypothetical_pnl_usd=199.11 > 0
(0 resolved alerts collapsing 0/0 → 0?)

That guard's own comment already names the diagnosis — "the tell of a 0/0 → 0
resolution collapse" — and that is precisely this window. It could not fire
before only because hypothetical_pnl_usd was also 0, from the upstream bug
that had the producer reading a warehouse without the derived tables
(auditmos/ogsfrompoly#234). Fixing that made the pathology detectable. The guard
worked as designed.

It cannot be satisfied honestly under schema_version: 1. No field separates
"0 of 0 resolved" from "0 of 240 resolved", so a truthful statement for an
unresolved window is unpublishable: either hit_rate: 0 asserts a 0 % success
rate that did not happen, or the PnL gets falsified back to 0 to match. Neither is
acceptable, and weakening the guard would throw away the one check that caught it.

Drafts are exempt from the guard and excluded by every consumer (excludeDrafts
in the index, the feeds and statements.md.ts; there is no per-statement route),
so draft: true removes the false claim from the live site immediately. The
corrected numbers stay in the file, making republication a one-line flip.

Root cause, for the record

hit_rate = in_favor / resolved, scored over alerts emitted in the window.
Measured on the producer host: 111 of 111 alerted markets join markets, 0
carry an outcome — against 97.3 % coverage warehouse-wide. Not the join, not
enrichment; the markets genuinely have not settled.

Every statement back to May carried 0.35–0.71, all tagged
[politics, macro-finance]. This is the first Macro-only statement — Politics
was retired 2026-07-24, and Politics markets (elections, daily events) supplied
essentially the whole resolved sample, while Macro markets resolve months out.
So this recurs every week until the schema can express it.

Follow-up that unblocks republishing

Add resolved_count, bump to schema_version: 2 across the producer and this
repo, and teach the sanity check that resolved_count === 0 makes hit_rate
vacuous rather than inconsistent. Then flip this file back to draft: false.
Tracked as auditmos/ogsfrompoly#236.

Verification

• Regenerated by poly-track report public weekly at auditmos/ogsfrompoly@554f537
  over the identical window
• Passes validate_statement (schema_version: 1)
• Verify (types, test, lint, knip) green
• No address leak — all wallet references are truncated IDs

Refs auditmos/ogsfrompoly#234, auditmos/ogsfrompoly#236