Parallel explore subagent fan-out dies to per-model 429s: explore's default model is the only rate-limited one, no backoff, no auto model switch despite eligibleForAutoSwitch

Describe the bug

Launching many subagents in parallel via the task tool concentrates all their model calls on one model bucket — explore agents all default to the same lightweight model (currently claude-haiku-4.5). That model appears to have a much tighter per-model burst limit than any other model, so a 16-agent explore fan-out hits HTTP 429 within ~20 seconds. Every subagent then fails with repeated integration_rate_limited errors and completes with empty output, while the parent session (on a different model) continues unaffected.

Three compounding problems:

1. The built-in explore agent defaults to the only model that rate-limits under fan-out. A survey of 1,089 local CLI sessions (~2.7 GB of events.jsonl) found exactly 5 sessions that ever logged errorType: rate_limit — every significant incident was a 16x claude-haiku-4.5 explore fan-out. Other models never triggered it despite far heavier use: gpt-5.6-sol (19k assistant messages, 201 sessions), claude-fable-5 (14k messages, including a clean 16-way fan-out re-run of the exact workload that failed on haiku), claude-sonnet-5 (9k), claude-opus-5 (7k), gpt-5.6-terra (5k) — zero incidents.
2. No backoff. The agentic loop retried into the same window at ~2 failing requests/second: 182 integration_rate_limited errors logged in 87 seconds in one incident.
3. No model fallback. The 429 response carries "eligibleForAutoSwitch": true, but subagents never switch — they just die and return empty results, silently wasting the whole fan-out.

Sample error from events.jsonl:

{"errorType":"rate_limit",
 "message":"You've hit the rate limit for this model. Please switch models or wait for your limit to reset in under a minute. Learn More (https://docs.github.com/copilot/concepts/rate-limits). (Request ID: C33A:2A98BA:2871A5:13FF8A5:6A7702DD)",
 "statusCode":429,
 "errorCode":"integration_rate_limited",
 "eligibleForAutoSwitch":true}

Observed concurrency data for the explore default model (same account, 4 incidents across 3 days):

• 16 parallel explore agents -> first 429 in ~20 s, all 16 stall and return empty
• 8 parallel -> clean
• ~15 launched sequentially over 10 minutes with overlap -> 2 transient errors, work completed
• 16 parallel on claude-fable-5 (model override) -> zero errors

Affected version

1.0.79-5 (Windows x64); incidents also observed on earlier 1.0.7x builds.

Steps to reproduce the behavior

1. In a session, use the task tool to launch ~16 explore subagents simultaneously (all default to the lightweight model).
2. Within ~20 seconds, subagent transcripts fill with "Limit reached — Resets in under a minute"; agents go idle after emitting only setup text and return empty results.
3. Re-run the identical fan-out with a model override to a larger model — it completes cleanly.

Expected behavior

Any (ideally all) of:

• Subagents honor eligibleForAutoSwitch and fall back to another model instead of dying (related: #2840).
• The agentic loop backs off per the reset window instead of retrying ~2x/s into the same limited minute (related: #2760).
• The CLI throttles subagent fan-out concurrency per model client-side, since it knows how many concurrent loops it is aiming at one bucket (related: #2545).
• The built-in explore agent''s default model either gets burst headroom matching the "fast, lightweight, fan out in parallel" positioning, or the CLI spreads large fan-outs across multiple eligible lightweight models.

Additional context

• The failure is invisible from the parent''s perspective until results come back empty: subagent.completed events fire normally with ~200-byte payloads.
• Nudging stalled subagents (write_agent) during the limited minute makes it worse — each nudge adds more 429s to the same window.
• Request IDs from multiple incidents available on request.

Commencez par reproduire l’échec avec task tool en utilisant un 16-agent explore fan-out, puis comparez events.jsonl pour le default model et un model override. Done devrait empêcher les retries répétés de 429 et garantir que les subagents renvoient des résultats utiles grâce au backoff, au throttling ou au model switching ; le payload ne nomme aucun fichier d’implémentation ni aucun test.