WhaleFlow swarm: synthesis/reduce pass (many workers → one coherent output)

[Hmbown]

Context

From the WhaleFlow vision-alignment assessment (docs/WHALEFLOW_VISION_ALIGNMENT.md). Both kimi-code's swarm and the ultracode pattern end with a synthesis stage that reduces many workers' outputs into one coherent result.

Problem

WhaleFlow has no reduce/synthesis pass. A fan-out produces N independent worker results with no merge step — the orchestrator does not combine, dedupe, or resolve conflicts across them.

Proposal

After a swarm/goal fan-out, run a synthesis pass — an orchestrator turn (or a dedicated synthesizer worker) that:

• consumes the structured worker-return envelopes (status / summary / artifacts / assumptions)
• reduces them into one coherent output
• dedupes overlapping findings and resolves conflicts
• surfaces a single result plus the contributing workers

Depends on

• The structured worker-return contract (v0.8.61: Add parent-visible worker interaction contract for agent/fleet fanout #3226 / Train 3 "Bug 3") — synthesis needs typed returns, not free-form chat.
• The WhaleFlow coordination substrate (companion issue) — synthesis reads the shared task ledger to know which results to reduce.

---

🤖 Filed from the WhaleFlow vision-alignment analysis.

[Hmbown]

Companion: #3229 (the coordination substrate this synthesis pass reads from).

[Hmbown]

The many-workers fan-out this synthesis/reduce pass depends on needs two substrate pieces, now filed for v0.8.63:

• v0.8.63: Queue-and-drain admission for high-fanout Workflows (decouple total Agents from the concurrency window) #3318 — admit-and-drain more Agents than the concurrency window instead of erroring at the cap (mod.rs:2300). Without this, a wide swarm can't even be enqueued.
• v0.8.63: Token-budget governor for Workflows / high-fanout Agent runs #3319 — shared token-budget governor so a wide fan-out paces spend (a quick test showed 20 trivial Agents ≈ 174k tokens in ~9s; swarm width needs a spend ceiling, not just a count cap).

Synthesis/reduce rides on top of those. Linking for sequencing.