fix(api): a submitted run is readable before it ends

[linhdmn]

POST /v1/runs registered nothing until the loop returned, so the run id it handed back resolved to 404 for the whole run. The documented submit-then-poll flow could not poll, and POST /v1/runs/{id}/kill (P75, NFR-1) had no run to signal — killing a live run answered run not found.

What changed

• submitRun registers the run before any exit path, so every exit writes its outcome through storeRunIfPresent instead of depending on a row that was never created.
• LoopRunner.WithProgress publishes a partial result at every step boundary (and once before the first step, and once after the loop), so a poll sees the steps that have run rather than nothing.
• All writes go through storeRunIfPresent, which refuses to create a row. DELETE removed the key and the loop's own final write re-inserted it, so a run an operator removed came back.

The bug found while fixing the bug

A blocked goal's 201 pointed at a run that was never stored at all — the goal-screen verdict was unreadable too, so the P1 guardrail path was invisible from the API. TestGuardrailScreenFiresInProduction caught it once storeRunIfPresent stopped creating rows.

Verification

go test ./... -count=1 green, gofmt clean, golangci-lint 0 issues, docs/check-prd.py OK. Both directions checked by deleting the registration line and watching TestRunVisibleWhileInFlight and TestGuardrailScreenFiresInProduction fail.

docs/PRD.md footer stamp updated in the same commit.