Add /story driver skill

[sebastientaggart]

Adds the /story driver skill (last child of story #167) and the small /start edit that lets it pass routine per-ticket prompts.

New skill skills/github-agile/story.md:

• /story <parent#> reads the parent's sub-issues, filters to open, and echoes the plan for one-time go approval.
• Loops sub-issues in GitHub-defined order. For each: pre-flight escalation check (ambiguous AC, sensitive surface, project-flagged domain) → invoke /start with the driver preamble → observe /submit-for-review tier routing.
• Per-ticket attempt cap of 2 and a no-progress guard (byte-identical diff between attempts) prevent unbounded retries.
• Posts a session-summary comment to the parent on completion. Sub-issue state is the source of truth — deleting or editing the summary does not break resumption.

/start edit (Case B prefatory block): When the immediately preceding context shows a [story-driver: parent=N ticket=K of M] preamble, three routine prompts are auto-passed: Step 2's plan-confirm gate, Step 3's investigation-findings prompt (default skip), and Step 5's verified-locally prompt. Every real escalation trigger remains; non-driver use of /start is unchanged.

What /story deliberately does not do (per the design discussion in #172):

• Does not auto-decide on /submit-for-review's needs-attention (WARNING) tier — operator decides address-now / follow-up / accept-as-is.
• Does not auto-decide on must-address (CRITICAL) tier — blockers still stop and wait for fixes.
• Does not write or edit code; only sequences /start and /submit-for-review.

The ambitious "automate judgment moments" path is deferred to #177, which may never be implemented depending on what day-to-day use reveals about where real friction lives.

Phase concept dropped. The original ticket mentioned phase boundaries with TBD on how phases would be defined. After discussion (label-list pollution, opinionated process, the "28 phases" stress test), we dropped the phase construct entirely. The driver walks all open sub-issues; the organic escalation triggers are the only stop points.

Untested in practice — first real use will be invoking /story on a real story. The orchestration design assumes the agent can recognize the driver preamble in the conversation context.

Part of #167.

Closes #172

[sebastientaggart]

Review Summary

Verdict: APPROVE

Findings

• [NOTE] PR is pure skill prompt documentation: adds the new /story driver skill source (skills/github-agile/story.md) and a small Case-B preamble-recognition block in skills/github-agile/start.md, with all four adapter outputs (.claude/, .agents/, .cursor/, .gemini/) regenerated. Running ./sync.py --force locally reports every skill up-to-date, so generated files are in sync with sources.
• [NOTE] Safety guardrails are explicit and conservative: attempt cap of 2 per sub-issue, byte-identical-diff no-progress guard, pre-flight escalation check for sensitive surfaces (auth, payments, secrets, prod config, destructive ops), and preserved human-judgment stops at /submit-for-review's needs-attention and must-address tiers. The /start Case-B edit only suppresses three named routine prompts and explicitly preserves all real escalation triggers.
• [NOTE] Sub-issue state remains the source of truth — the Step 5 summary comment is documented as informational so deletion/editing does not break resumption, matching the PR description's intent.