[DEBT] Make execute session prewarm lazy (Lever F1)

[AdrienneBosch]

Context

In a real /github-workflow:execute run (2026-06-24), the Session
prewarm fired three gh calls — candidates, label inventory,
rate_limit — before the Phase 1 fast path (wf pick) returned ok.
The execute body's own note states .claude/candidates.json "is never
read" when the fast path succeeds. So on the happy path (the path that
actually ran), the candidate fetch was provably wasted, and the
~120-line inline fallback it was warming never executed.

The prewarm predates wf owning pick→claim→board→branch in one call. It
eagerly front-loads work for the inline fallback, inverting priorities:
the common path pays to warm the rare path.

See docs/context-optimization-plan.md → Lever F1. This is the
lowest-risk lever: it removes nothing — the inline fallback stays fully
intact for the Python-less / wf-broke case. It only stops pre-paying
for a path that usually never runs.

Requirements (acceptance criteria)

• The candidate fetch (.claude/candidates.json) is gated on the
  fast path NOT returning ok — it runs only when wf pick yields
  unsupported/error/missing-interpreter and the inline fallback is
  actually entered.
• Trace where .claude/label-cache.json is consumed (board moves now
  done by wf; finish phase). If no happy-path phase reads it, gate the
  label-inventory fetch the same way; if a later phase still needs it,
  document which and keep it (or defer to first use). Record the finding in
  the PR.
• The eager rate check stays (cheap, keep-in-context).
• The prewarm prose moves into a references/ file loaded only when
  the inline fallback fires (one level deep from SKILL.md).
• The inline fallback still works unchanged when entered (no capability
  removed; Python never becomes a hard dependency).
• Quality gate green; offline tests green; one real end-to-end dry run
  on both the ok fast path and a forced-fallback path.
• _shared-skills/ source edited if shared, sync run, versions bumped.

Savings goal

Eliminate 2–3 wasted gh API calls per happy-path run (latency +
rate-limit budget) and remove the ~58-line prewarm section from the hot
path. Target: zero eager candidate/label fetches when wf pick returns
ok.