knowledge/youtube-digest: persist the resolved --target in WatchState so resume can recover it

[kyle-sexton]

Deferred from the Codex review on #1027 (P2, SKILL.md L289). #1027 closed the
--target local-checkout resolution gap at the contract layer and deliberately kept the
machine-local checkout path out of the staged slice README.md; persisting the resolved
target through the extraction runtime is the remaining half and is a code change, out
of that PR's documentation scope.

Finding

watch --target <repo> resolves a synthesis target, but nothing in the extraction runtime
persists it:

• extraction/watch/watch-state.js — the WatchState typedef (fields: videoId,
  videoSlug, sourceUrl, title, status, phases, skipResearch, frameSelection,
  artifactPaths, tempSession) has no target field, and createWatchState() writes
  none.
• buildContinuationPrompt() (same file) tells a resumed session to read
  run-state/watch.json, source/transcript.txt, and the lane deliverables — never the
  slice root README.md that carries the **Target:** line.
• extraction/watch/run-resume.js reads only run-state/watch.json.

So a cross-repo watch --target <repo> interrupted before synthesis loses the resolved
target: resume must re-infer or re-ask.

Fix direction

Persist the resolved target name in WatchState when the watch starts, and surface it
in buildContinuationPrompt() so the resumed session re-runs the SKILL.md "Synthesis
target resolution" rungs against the recorded name instead of re-asking.

Portability constraint (the reason #1027 did not just write the path anywhere): the
resolved local checkout path is machine-local. run-state/watch.json is Staged: yes per
the Output contract and is already {tmp}-tokenized by serializeTempSession() /
sanitize-slice-temp-paths.js — so persist the portable name, not the absolute tree
path, or route the path through the same tokenization/sanitization seam.

Acceptance

resume on a watch started with --target <repo> recovers the target without re-asking,
and no staged artifact gains a machine-local absolute path.

Related

• docs(knowledge): fix youtube-digest contact-sheet retention wording + --target resolution gap #1027 — PR that closed the resolution half and raised this deferral.
• knowledge/youtube-digest: SKILL.md contact-sheet retention wording + --target local-checkout resolution gap (PR #854 follow-ups) #1015 — parent batch (Site 2, --target local-checkout resolution gap).
• docs(knowledge): youtube-digest synthesis target + named agnosticism gaps #854 — originating PR whose review first deferred the target-durability nit.

[kyle-sexton]

This was generated by AI during triage.

Triage

Verified. Confirmed in extraction/watch/watch-state.js: the WatchState typedef (videoId, videoSlug, sourceUrl, title, status, phases, skipResearch, frameSelection, artifactPaths, tempSession) has no target field, and createWatchState({ videoId, videoSlug, sourceUrl, title }) writes none. buildContinuationPrompt() in the same file directs a resumed session to run-state/watch.json, source/transcript.txt, and lane deliverables — never the slice-root README.md **Target:** line. Confirmed run-resume.js reads only run-state/watch.json. The gap is exactly as described: a cross-repo watch --target <repo> interrupted before synthesis loses the resolved target on resume.

Classification: Task (deferred code-review nit, not a reported production failure), priority: low, work-class: scoped.

Agent Brief

Type: Task
Summary: Persist the resolved synthesis-target name in WatchState so a resumed watch session recovers it instead of re-asking, without leaking a machine-local absolute path into staged artifacts.

Current behavior: watch --target <repo> resolves a synthesis target but the extraction runtime never records it. WatchState has no target field, and the resume continuation prompt never points a resumed session at re-deriving it.

Desired behavior: The resolved target's portable name (not the absolute local checkout path) is written into WatchState when the watch starts, and buildContinuationPrompt() surfaces it so a resumed session re-runs the target-resolution logic against the recorded name instead of re-asking the user.

Key interfaces:

• WatchState typedef — needs a target-name field alongside the existing videoId/videoSlug/sourceUrl fields.
• createWatchState() — needs to accept and persist the resolved target name when one is present.
• buildContinuationPrompt() — needs to surface the recorded target name in its resume guidance.

Acceptance criteria:

• resume on a watch started with --target <repo> recovers the target without re-asking.
• No staged artifact (anything marked Staged: yes in the output contract) gains a machine-local absolute path — the persisted value is a portable name, or is routed through the existing temp-path tokenization/sanitization seam.

Out of scope:

• Any change to how the target is initially resolved (that gap was already closed by docs(knowledge): fix youtube-digest contact-sheet retention wording + --target resolution gap #1027) — this only concerns persistence and resume recovery.

[kyle-sexton]