Replace the bars task's phantom retry with a repeating trigger

[mspinola]

What

WINDOWS_SCHEDULING.md built its whole retry story on Task Scheduler's "if the task fails, restart every N minutes." That setting fires when the scheduler cannot launch the action — not when the action exits non-zero — so it never once retried a --require-final defer.

Evidence

Measured on the producer box, which had RestartCount 20 / RestartInterval PT15M set on the bars task:

Date	Action exit	Launches that night
2026-08-12	1 (defer)	1
2026-08-13	1 (defer)	1
2026-08-14	1 (defer)	1
2026-08-15	1 (defer)	1

Four consecutive nights, no retry, no bars captured. Events 111 and 322–324 — the restart and queue events — never appeared at all. The store caught up on 08-16, when a run finally found Norgate ahead of it.

That self-heal is why it stayed invisible: the data was never permanently wrong, only a day late, so --check a week later looked fine and nothing alerted.

Fix

A repetition on the trigger, which fires regardless of what the previous run returned. schtasks sets it on a daily schedule with /RI and /DU, verified to emit the same <Repetition> element the Friday COT poller has always used — which is why that task worked all along while the bars task did not. The correct pattern was already on the page; the bars task just wasn't on it.

Also corrected, downstream of the same premise

• Last Result guidance now says the opposite of what it did. Under repetition a healthy night ends non-zero, because the repeat that captured succeeded hours earlier and every repeat after it defers. A red Last Result on the bars task is no longer evidence of anything on its own.
• The exit-code callout for run-prices.cmd no longer claims an unguarded --metadata strands the task until tomorrow. It doesn't any more — what it destroys is the only per-run signal separating a capture from a defer.
• sync-store.cmd's exit-code normalisation rationale no longer cites a restart loop that doesn't exist.

The warning keeps the dates and event IDs rather than quietly swapping the recommendation, because the failure mode is invisible enough that someone will otherwise re-derive it from scratch.

Applied to the producer box

• cotdata prices converted to a repeating trigger (15 min × 5 h); verified firing on exit 1.
• cotdata COT (catch-up) had its dead RestartOnFailure element removed. It needs no polling — cotdata-cot --cot-all exits 0 on a pre-release no-op — and being daily and idempotent, the next morning's run is its own retry.

Note for reviewers

The unmerged branch claude/scheduling-doc-finals-gate touches this same file and looks superseded: its finals-gate rewrite already landed in main in evolved form, and merging it now would re-introduce the restart-on-failure paragraph this PR removes, plus pre-ADR-0007 cotdata-prices --prices commands. Worth closing rather than merging.

🤖 Generated with Claude Code