Addresses review feedback on PR anomalyco#30482: AGENTS.md prefers inlining
single-use helpers. The prior commit (e97a3d2) already removed
sapGeminiReasoningParams (folded into the shared googleThinkingVariants
helper) and sapOpenaiReasoningParams (inlined). This drops the last
SAP-specific single-use helper.

Multi-use helpers are kept: wrapInSapModelParams (3 call sites) and
googleThinkingVariants (2 call sites, shared between native and SAP).

…r SAP models

- Gate Gemini branch on `2.5` so non-2.5 (Gemini 3+) falls through to
  the harmonized fallback. SAP vertexai harmonizer drops unknown
  thinkingLevel; harmonized reasoning_effort is mapped server-side.
- New harmonized fallback at the end of the SAP case using
  `reasoning_effort: low|medium|high` for any reasoning-capable model
  not handled by the major-provider branches (cohere reasoning,
  sonar-deep-research, future providers).
- Update tests: gemini-3.1-flash-lite now asserts harmonized fallback;
  add cohere-command-a-reasoning and sonar-deep-research fallback tests;
  llama opus-substring test asserts harmonized fallback (previously
  asserted empty).

Consolidate Gemini 2.5 pro/flash tests into one loop parametrized on
maxBudget. Consolidate gpt-5/-mini/-nano + gpt-5.4 + o3-mini into one
loop parametrized on apiId+releaseDate+efforts. Consolidate the four
harmonized-fallback tests (gemini-3.1-flash-lite + cohere + sonar +
llama opus-substring) into one loop. Net -16 lines, same coverage.

- Drop redundant .toLowerCase() on the `2.5` check in the Gemini
  guard; matches the case-sensitive substring style used by the
  surrounding anthropic/gpt branches (SAP API ids are already
  lowercase).
- Parametrize the non-adaptive Anthropic test over claude-sonnet-4
  and claude-4.5-opus to lock in that pre-4.6 reasoning models still
  take the budget_tokens path.

Drops the local lowercased id alias. The googleThinkingBudgetMax and
googleThinkingLevelEfforts helpers already lowercase their input, and
the inline "2.5" substring check matches against lowercase Gemini ids
by convention. Single-source naming aligns with the surrounding switch
which uses model.api.id directly.

Aligns the googleThinkingBudgetMax call with the surrounding
`id.includes("2.5")` check (outer-scope `id = model.id.toLowerCase()`)
introduced by the prior commit's drive-by fix.

Preserves the defensive .toLowerCase() the original @ai-sdk/google[/-vertex]
branch had on the outer scope id (= model.id.toLowerCase()). The helper
now declares const id = model.api.id.toLowerCase() locally and reuses it
for every check and every helper call inside googleThinkingVariants.

Replaces model.api.id with the outer-scope id (= model.id.toLowerCase()
declared at L643) in the SAP case, matching the majority pattern used by
other case branches (gateway, anthropic, copilot, etc.). The helpers
GPT5_FAMILY_RE and openaiReasoningEfforts are anchored on (?:^|/) so they
match equally well against the prefixed model.id form.

Mirrors the .serena/ pattern for the .omo/ agent runtime directory.