feat: one-time GitHub star prompt at Stand Down

[harrymunro]

Summary

• Adds a single, one-time GitHub star prompt to Step 8 (Stand Down) that fires only on a successful mission, after the Mission Complete Gate.
• Pattern mirrors Q00/ouroboros: a pure markdown directive in SKILL.md — no new Python module, subcommand, event type, or test surface.
• Persists {"star_asked": true} in ~/.nelson/prefs.json so the prompt is one-shot per user across all Nelson projects on the same machine.

Behaviour

The directive tells Claude to:

1. Skip silently if gh auth status fails or ~/.nelson/prefs.json["star_asked"] is already true.
2. Otherwise invoke AskUserQuestion with two options: Star Nelson / Maybe later.
3. On Star Nelson: run gh api -X PUT /user/starred/harrymunro/nelson (idempotent — no pre-check needed). On API failure, print a friendly note and continue; never block Stand Down.
4. On any answer (including a custom "Other" response): merge star_asked: true into prefs via inline python3 -<<'PY' heredoc, preserving any existing keys.

Out of scope (deferred per issue notes)

• Re-asking after N missions on "Maybe later".
• Configurable NELSON_NO_STAR_PROMPT env var (manual workaround documented in README).
• /nelson-star slash command.
• Mission-log telemetry beyond the local prefs flag.

Test plan

• scripts/check-references.sh passes (no broken cross-references).
• pytest skills/nelson/scripts/ — 278 tests pass.
• pytest hooks/ — 52 tests pass.
• Heredoc smoke-tested against a tmp HOME: fresh-create, key-preservation merge, read-back idempotency, missing-file branch all behave correctly.
• Manual QA on a real mission: gh authed + first success → prompt fires, Star Nelson → repo starred, prefs file written.
• Manual QA: second mission Stand Down → prompt does not re-fire.
• Manual QA: gh auth logout → prompt skipped, Stand Down completes.
• Manual QA: outcome_achieved=false → prompt skipped.

Closes nelson-4yt.

[harrymunro]

Reviewed and addressed in bf20cbc.

Changes

1. Defensive prefs reader — The read-side heredoc now mirrors the writer's shape: missing file, malformed JSON, and non-dict top-level all fall through to PREFS_OK instead of raising. A corrupted ~/.nelson/prefs.json will now skip the prompt silently rather than surface a Python traceback in chat.
2. Disk-based outcome gate — Added a third preflight heredoc that reads {mission-dir}/stand-down.json and emits OUTCOME_OK / SKIP_OUTCOME_NOT_ACHIEVED / SKIP_NO_STAND_DOWN. The prompt no longer depends on the model remembering whether --outcome-achieved was passed minutes earlier — it's an observable signal on disk. Driven by MISSION_DIR env var with a {mission-dir} placeholder fallback.
3. Canonical slug — Directive now names harrymunro/nelson once up front as the canonical slug. The two gh api references remain because they're command lines the model runs verbatim.
4. README placement — Explainer paragraph moved below the star-history image under a new ### How the star prompt works subhead so the section still leads with the chart.

Verification

• All 8 reader/outcome branches smoke-tested against tmp dirs (missing/malformed/non-dict/already-asked × outcome=true/false/missing/malformed) — every signal lands as documented.
• pytest skills/nelson/scripts/ → 278 passed.
• pytest hooks/ → 52 passed.
• scripts/check-references.sh → OK.