Add WorkBuddy desktop setup recipe and 2026-09-12 thinking deviations

[JJasonSun]

Summary

Adds the WorkBuddy desktop custom-model configuration recipe and four new dated
deviations from an owner-authorized interactive integration session
(2026-09-12 through 2026-09-14, Windows x64, Node.js fetch direct HTTP).

New file

• references/workbuddy_setup.md — Field-by-field WorkBuddy models.json
  configuration for ECNU models, including:
  • The bare-array format requirement (the desktop local-model service purges
    object-shaped files to [] on every startup when the hardware gate fails).
  • Why the Settings editor must not be used to save (it uses parseModelsJson
    which does not accept arrays).
  • Effort field semantics (supportedEfforts, defaultEffort, canDisableThinking).
  • How effort reaches the API (priority chain + identity thinkingLevelMap).
  • Why thinking is never in the request (catalog miss is harmless for ECNU).
  • Runtime verification via usage.completion_tokens_details.reasoning_tokens.

New dated deviations (environment live-2026-09-12-c)

Deviation	Key finding
ecnu-reasoner alias default thinking	Bare alias requests activate thinking server-side; no client-side thinking parameter needed.
ecnu-max reasoning_effort as thinking trigger	reasoning_effort alone activates thinking on ecnu-max, contradicting the documented gating.
Unavailable reasoning effort tiers	minimal and medium return intermittent HTTP 500; xhigh is undocumented but stable.
Rapid-request 401 metadata failure	Sub-second sequential requests return 401 mimicking auth failure; ~4s spacing eliminates it.

Updated routing tables

SKILL.md, README.md, AGENTS.md, workflows.md, examples.md, and
models.md now cross-reference the new recipe and deviations.

Validation

• python3 scripts/validate_skill.py → passed
• python3 -m pytest tests/ -v → 69 passed, 1 skipped, 57 subtests passed
• Secret scan (sk-[A-Za-z0-9_-]{16,}) → no matches
• Personal path scan (/Users/, C:\Users\, /home/) → no matches
• All deviation entries include the 8 required fields and use valid statuses
  (active).

Notes

• The dose-response medians (low ≈117, high ≈155, xhigh ≈178, max ≈343
  reasoning tokens) are characterization samples, not quality benchmarks.
• No SDK, streaming, long-context, embedding, rerank, TTS, or image-generation
  revalidation was performed in this session.
• Older observations retain their original dates and statuses.