fix(llm-to-bedrock): sweep stale Sonnet 4 example model IDs to Sonnet 4.6

[herosjourney]

Summary

The llm2bedrock agent examples (code-rewriter, code-analyzer, report-generator) and the behavior-delta reference still used us.anthropic.claude-sonnet-4-20250514-v1:0 — Claude Sonnet 4 (May 2025) — as the example rewrite target. The plugin's own model catalog marks that model excluded, EOL Oct 14, 2026 (ai-model-lifecycle.md), and every recommendation elsewhere points at Sonnet 4.6. An agent pattern-matching the examples would produce rewrites targeting a model the plugin itself bans. This PR sweeps the examples to claude-sonnet-4-6 and adds a guard note so example IDs can't silently drive rewrites again.

Changes

• Swept claude-sonnet-4-20250514-v1:0 → claude-sonnet-4-6 in:
  • agents/llm2bedrock-code-rewriter.md (7 sites: Vercel AI SDK / LangChain / boto3 examples, env-var block)
  • agents/llm2bedrock-code-analyzer.md (target_models example)
  • agents/llm2bedrock-report-generator.md (model-id list example, pricing comment — Sonnet 4.6 has identical $3/$15 rates so the comment's math is unchanged, dual-ARN prose example)
  • skills/llm-to-bedrock/references/helpers/behavior-delta-detection/references/openai-to-bedrock.md (converse example)
  • skills/llm-to-bedrock/scripts/test_iam_policy.py, test_validate_result.py (example IDs in tests; inference-profile logic is ID-agnostic)
• Intentionally kept: the Sonnet 4 rows in pricing-cache.md and ai-model-lifecycle.md — those are catalog entries whose whole point is recording the model and its EOL/excluded status.
• New guard note in llm2bedrock-code-rewriter.md above the rewrite examples: example IDs are illustrative; authoritative IDs come from the migration plan (aws_model_id); verify plan IDs resolve via aws bedrock list-foundation-models (or the preflight converse-ping); if an ID no longer resolves, STOP and surface it rather than substituting a model. This mirrors the agent-toolkit-for-aws guidance on verifying model IDs at runtime.

Relationship to other PRs

Independent — based on current main (post-#162). No overlap with open PRs (#141/#144/#149/#151/#155/#156/#161/#163).

Test plan

• Repo-wide sweep verified: zero claude-sonnet-4-20250514 references outside the two catalog files
• pytest on the two touched test scripts: 119 passed
• Full mise run build green (160/160 checks)
• Fresh-agent rewrite replay against the fixture repo (examples-only change; flagging honestly that no replay was run)

[herosjourney]

Applied the Cursor review (commit 89eda24) — all five points, with two findings that came out of verifying point 2:

1. Regression lint — added. tools/model-id-lint.py (stdlib) + lint:model-ids mise task in the lint group, so it runs in every build. Two pattern classes: the EOL Sonnet 4 ID outside the two catalog files, and fabricated dated forms (sonnet-4-6-<date> / opus-4-8-<date>). Explicit per-pattern allowlists; extend BAD_PATTERNS on future EOLs.

2. The "dated Sonnet 4.6" IDs are fabricated, and one had a wrong price. claude-sonnet-4-6-20250514-v1:0 grafts Sonnet 4.6's name onto Sonnet 4's date stamp — that ID never existed. Swept in the three gcp design refs. Two things Cursor's list didn't surface:

• bedrock_pricing.py's STATIC_FALLBACK carried the fabricated dated forms as executable lookup keys, plus a fabricated claude-opus-4-8-20250610 at $15/$75 per 1M — a 3x overstatement (Opus 4.8 is $5/$25 per the plugin's own pricing cache). Table now uses real undated IDs with the correct rate; _static_fallback strips date stamps so dated inputs still resolve; two new tests guard the rate and the date-strip (130 pass in the touched files; the 3 pre-existing test_preflight_bedrock.py failures reproduce on untouched main — env-only, missing boto3).
• resolve-bedrock-model-id.md keeps its dated example on purpose (repairing broken IDs is that helper's job) — now labeled intentionally-invalid, and allowlisted in the lint.

Swept in this PR rather than a separate one: the lint has to be green to merge, so lint + sweep are one atomic change.

3. Analyzer guard — added as a _comment_target_models sibling key in the example JSON.

4. ID-form note — added to the rewriter guard: geo-prefixed inference-profile form preferred in rewritten code; bare foundation-model IDs belong in pricing tables (they often reject on-demand invocation — see the IAM known-fix).

5. Replay — left unchecked, per the review.

Full mise run build green (160/160) including the new lint.

[leon1418]

Currently Bedrock supports GPT-5.6, and Sonnet is positioned in the same tier as GPT-5.6 Terra. So, should we recommend migrating to a model from the same providor?