fix(sse): mark gemini-3.5-flash as thinking-capable

[diegosouzapw]

Closes #10286

Root cause

src/shared/constants/modelSpecs.ts declares the shared GEMINI_35_FLASH_MODEL_SPEC
constant with supportsThinking: false. The base "gemini-3.5-flash" entry spreads
that spec, so getResolvedModelCapabilities() reports the base Google AI Studio model
as thinking-incapable, which makes capabilityFor()
(src/lib/reasoningRouting/policy.ts) resolve to "unsupported" for any
reasoning_effort, and applyDecision() (src/sse/handlers/reasoningRouting.ts)
turns that into a pre-provider HTTP 400 — even though the base model supports
reasoning (it has an effort-tier alias gemini-3.5-flash-high, and the reporter
independently verified flipping the flag to true works end-to-end).

The same shared GEMINI_35_FLASH_MODEL_SPEC constant is also spread into several
Antigravity flash-tier aliases (gemini-3.5-flash-low, gemini-3.5-flash-extra-low,
gemini-3-flash-agent, gemini-3.6-flash-*), and those tier ids are documented as
rejecting client-supplied thinking params (the model id itself selects the reasoning
tier upstream). So the fix does not flip the shared constant — it sets
supportsThinking: true as an explicit override on the base "gemini-3.5-flash"
entry only, leaving GEMINI_35_FLASH_MODEL_SPEC and all the Antigravity tier aliases
unchanged (still false).

Fix

src/shared/constants/modelSpecs.ts: added supportsThinking: true override on the
base "gemini-3.5-flash" entry (after the ...GEMINI_35_FLASH_MODEL_SPEC spread).

Regression test

tests/unit/gemini-3-5-flash-thinking.test.ts — TDD, fails on unfixed code
(AssertionError: false !== true / decision.capability 'unsupported' vs
'supported'), passes after the fix:

✔ gemini-3.5-flash (AI Studio provider) resolves as thinking-capable
✔ reasoning_effort 'high' on gemini-3.5-flash is NOT rejected by routing policy

Asserts both:

1. getResolvedModelCapabilities({provider:"gemini", model:"gemini-3.5-flash"}).supportsThinking === true
2. reasoning-routing policy resolves reasoning_effort:"high" on gemini-3.5-flash to
   capability "supported" (not the pre-provider 400).

Gates run

• npm run typecheck:core — exit 0
• npx eslint --suppressions-location config/quality/eslint-suppressions.json <changed files> — clean
• node scripts/check/check-file-size.mjs — no new violation on modelSpecs.ts
• node scripts/check/check-complexity.mjs — OK (2456 violations, baseline 2774)
• node scripts/check/check-cognitive-complexity.mjs — OK (1104 violations, baseline 1223)
• node scripts/check/check-test-discovery.mjs — OK, new test file discovered
• npm run test:unit — started full suite; 614/4565 test files completed with 0
  failures before the run stalled to near-zero throughput under severe 13-way
  parallel devbox contention (13 concurrent fan-out agents running the identical
  full suite on shared hardware — not specific to this change). No regression
  signal in the portion that ran. CI runs on isolated runners and is the
  authoritative full-suite gate for this PR.

⚠️ base-red inherited: #9985 — ESLint errors (2) from #10250