Strip scheduler restatements from the Skywalker prompt

[TheGreatAxios]

Summary

• The Skywalker system prompt no longer restates runtime scheduler mechanisms; each deleted section names behavior the runtime already enforces in code, so the prompt keeps orchestration judgment and drops mechanism copy
• Prompt shrinks by ~1.8k characters with no behavior change on any surface; every remaining rule is judgment or taste, and no numeric limit survives that the runtime does not enforce

Verification

• bun run typecheck passes (exit 0)
• bun run lint passes (oxfmt + oxlint, 0 warnings, 0 errors)
• bun run build passes
• bun test ./src ./tests ./evals ./scripts --randomize --seed 424242 passes (7131 pass, 0 fail, 488 files)
• Per-row audit against current main — landed means deleted, with the mechanism evidence in code:
  • Admission queue (hard caps): landed — src/subagent/admission.ts:137,194 (queueing + width caps) — deleted queues excess restatement
  • Overlap detection (lane disjointness): landed — src/subagent/admission.ts:64 (laneDisjointness) — deleted queue-depth restatement
  • Composable stop conditions (retry budgets): not landed as a runtime budget — no cap exists in code, so the numeric Cap re-fix rounds (e.g. 1-2) is deleted while the loop judgment (ship -> verify -> fix -> re-verify) is kept
  • Steering + resumable sessions: landed — src/subagent/agent-fleet.ts:510 (spawn returns immediately), src/subagent/exec/runner.ts:144, src/agent/agent-search.ts:114 — deleted mount-fact and triple-stated contract restatements; kept the fail/successor/interrupt/cancel judgment
  • Non-blocking spawn: landed — same three files — deleted the static mount fact; the unmounted chat prompt loses an impossible-tool ref (CL-6807 hygiene) while mount-correct copy stays in the mount-gated guidelines and tool descriptions
• Kept remainder: classification, routing table, effort scaling as judgment, report tone, and the fail-then-successor / interrupt-resume / operator-cancel-wait judgment
• Shared sections with sibling CL-7665 (src/agent/directors/skywalker/package.ts, markers): that branch compresses classification while this one strips scheduler prose — the second to land rebases

Fixes CL-6953

[linear-code]

CL-6953

[TheGreatAxios]

Audit: Needs-work, fix now pushed as 1c93834. Loop repro: the branch deleted "Cap is one successor for that stall" while keeping MAY spawn_agent one successor with a changed brief per stall. Each stall permits one successor but nothing terminates the chain: the successor stalls, its stall permits another one successor, ad infinitum, with each step plausibly changed enough to pass the identical-re-dispatch guard. Fix restores the qualitative bound (one successor per stall, then synthesize + Blockers and stop), pinned by expect(p).toContain("Cap is one successor for that stall"). Checks on the branch: bun run typecheck exit 0; bun test skywalker + prompts suites 44 pass, 0 fail.