feat(prd-intake): operator PRD checkboxes become loop work (the value lane #355 asks for)

[linhdmn]

Problem

The self-build loop ships loop plumbing because its deterministic lane is empty (#355), and there is no path at all from what the operator writes in docs/PRD.md to implemented work. Today a human PRD edit either (a) blocks the loop ([prd-fresh] … operator mid-edit, skipping iteration, internal/loopdriver/run.go:232) or (b) is ignored: the only PRD→work parser (pipeline.ParseBacklogItems) is reachable only from devagent task --pick, which the loop never calls (internal/loopdriver/dispatch.go:474 passes --prompt).

Requested change

A PRD intake step: an operator-authored open checkbox in docs/PRD.md becomes a queue row the next iteration claims.

• internal/prdintake: parse - [ ] … items (heading context + indented sub-bullets as acceptance criteria), skipping fenced code blocks and struck/completed lines; deterministic id PRD-<sha8> so re-runs dedupe against the existing queue store.
• Goal text must satisfy the dispatch contract (Goal: prefix, ≤120 words, internal/loopdriver/dispatch.go:450); the section body rides along as the queue task-PRD sidecar so the worker reads the real spec, not just the bullet.
• devagent prd-intake [--dry-run] [--json] [--max N] [--issues] prints what it queued.
• Loop wiring: intake runs at iteration head (after the PRD-currency gate, before claimQueueTask), SELFBUILD_PRD_INTAKE=0 disables.
• Dispatch rider tells the worker to tick the source checkbox to - [x] in the same PR, so a shipped item never re-enters the lane.

Acceptance

• A new - [ ] bullet committed to docs/PRD.md is enqueued and claimed by the next devagent loop iteration; running intake twice enqueues once.
• - [x], struck, and code-fence examples are never queued.
• Unit tests for the parser (context, nesting, fences, idempotent ids) and the ingest dedupe.

[linhdmn]

Implemented in the working tree 2026-09-14 (labels adjusted so the loop does not re-pick it before this lands).

• internal/prdintake + internal/loopdriver/intake.go + devagent prd-intake + internal/commands/updown.go (up/down) + internal/scout/run.go (live FR-SCOUT-01 cycle).
• Verified: gofmt/go vet/golangci-lint clean, full go test ./... green, and live smoke runs — an open - [ ] item in a scratch repo's docs/PRD.md was ingested, claimed by the next iteration, and published as .selfbuild/goals/loop-1.md; up proved the detached driver (holds the loop lock, iteration 2, phase queue-empty) and failed it with the cause (exited: max iterations reached) when the driver halted; down stopped the recorded pid and left no process behind; devagent scout --once --dry-run previews a cycle with no AI call and no writes.