feat(deepseek): flip operator default to v4-flash (post-eval)

[michaeljboscia]

Summary

Lands the default-model flip that PR #39 held back pending empirical eval. Data is in daemon/docs/v1-deepseek/PRO_VS_FLASH_EVAL_RESULTS.md.

• Default cfg.model changes from deepseek-v4-pro → deepseek-v4-flash.
• Per-call AskAgentRequest.deepseek_model override and TRIUMVIRATE_DEEPSEEK_MODEL=deepseek-v4-pro operator escape hatch both preserved.
• Workspace: 292/292 unit tests pass.

Empirical case (130 consults / 13 task types over 4 eval waves)

Wave	Sample	Pro pass	Flash pass
v1	6 × 1	6/6*	6/6*
v3	9 × 5	44/45	44/45
v4 (Pro-favoring)	4 × 5	19/20	19/20

v4 specifically probed the four categories where published DeepSeek benchmarks claim a material Pro advantage (long-context cross-file synthesis, complex 15-field nested JSON, hard algorithm with space optimisation, SimpleQA-Verified-style factual recall). Quality tied. The single Pro failure was Pro emitting two valid JSONs concatenated with leaked meta-reasoning prose between them — i.e., breaking the explicit "Output ONLY the JSON object" instruction. The single Flash failure was an assertion-regex mismatch on a correct algorithm output. The only quality signal that emerged across all 130 consults cut against Pro, not for it.

Token economics: Pro used 8.6% more total tokens, 12.4% more reasoning tokens, at indistinguishable answer quality.
Cost (v4 promo pricing): Pro $0.053 vs Flash $0.016 for 20 calls each — ~3.4× more for Pro. Post-promo on 2026-06-01 the gap widens.

Honest test boundaries (called out explicitly in EVAL_RESULTS.md)

This eval is strong evidence for the daemon's current single-shot, bounded-context workload. It does NOT cover:

• Multi-turn agentic tool-use loops (Pro's Terminal-Bench edge: 67.9 vs 56.9)
• 100K+ token prompts with deep-needle retrieval (Pro's 1M-context edge)
• Judge-LLM quality grading (only binary contains/regex assertions used)
• Reasoning-effort sweeps (both ran thinking: enabled, effort: high)

If/when the daemon adds multi-turn flows or routes 100K-token prompts, re-evaluate. Both escape hatches above remain available immediately.

Files changed

• daemon/crates/mcp-bridge/src/deepseek_config.rs — default value + doc-comment + from_env_returns_documented_defaults_for_all_15_knobs test
• daemon/crates/shared-types/src/lib.rs — doc-comment on AskAgentRequest.deepseek_model
• daemon/docs/deepseek-operator-runbook.md — per-call + env-var tables
• daemon/docs/v1-deepseek/PRO_VS_FLASH_EVAL_RESULTS.md (new) — full results
• daemon/docs/v1-deepseek/PRO_VS_FLASH_DRIVER_SPEC.md (new) — co-located driver spec referenced by EVAL_RESULTS

Test plan

• cargo test -p mcp-bridge --lib — 106/106 pass
• cargo test --workspace --lib — 292/292 pass
• Manual smoke: mcp__triumvirate__ask_agent { agent: "deepseek", message: "..." } post-merge should route to Flash by default
• Manual smoke: same with deepseek_model: "deepseek-v4-pro" per-call should still route to Pro

🤖 Generated with Claude Code