[aheritier]

Part of #4654 (PR 1 of 3). Part of #4583 and epic #4580.

What

Three small tests pin Claude 5.x lifecycle behaviors that already work. Test-only: no production code changes, and no new test harness. Each test reuses what exists, per #4654's "test cost matters" rule.

Test	Pins	Runtime
TestEmbeddedSnapshotHaiku55ContextTier (pkg/modelsdev)	Haiku 5.5's price tier in the embedded catalogue: base rates up to 100,000 prompt tokens, tier rates above (99,999 / 100,000 / 100,001), for anthropic, which also prices Vertex, and for amazon-bedrock global and us. (+10%)	0.03s
TestRoutedPricing/catalog_leaf_survives_reload (pkg/runtime)	After a SQLite save and reload, a routed turn keeps the serving model's ID (not the router's) and its cost	0.02s
TestSharedClientConcurrentSessions (pkg/model/provider/anthropic)	Two concurrent sessions on one client don't share system prompt, user text, attachments or tools; each is sent exactly once, as itself	0.01s, run 20× under -race

Each test was checked by deliberately breaking the behavior it pins and confirming it fails:

• making the 100k threshold inclusive;
• excluding Cost from the stored JSON;
• leaking the system prompt through a package variable;
• replaying one session's whole request to the other.

Not added, and why

• "Real client to cost": the whole-request tier with cache reads and writes counted is already covered with synthetic rates (TestComputeMessageCostContextTiers). Anthropic's usage mapping, cache tokens included, is covered by TestStreamUsagePreservesCumulativeCounts.
• Bedrock shared client: the client holds no per-conversation state (SDK client and a caching flag; the block recorder lives per stream). Testing it at runtime would need a Bedrock fake server, which test(runtime): Claude 5.x lifecycle — compaction, reload, switches, shared clients #4654 rules out. Not established at runtime level.
• Vertex pricing: Vertex Claude is priced from the Claude API catalogue as a best estimate; regional surcharges are not modelled (test(runtime): Claude 5.x lifecycle — compaction, reload, switches, shared clients #4654).

The end-to-end runtime harness moves to PR 2, where compaction needs it.

References

Change	Documentation
Haiku 5.5: "$0.10 / MTok for prompts up to 100,000 tokens", "$0.50 / MTok for prompts over 100,000 tokens"; cache reads and writes priced at matching rates	Claude pricing
One client serves every conversation of a team load; per-conversation state belongs on the request	AGENTS.md, "Client and Session Lifetimes"

Validation

• task build passes.
• task test passes. On a first run, e2e TestA2AServer_MultipleRequests failed (404 instead of 200); it passed 3/3 alone and on a full rerun. That's flaky under load, and this PR doesn't touch it.
• golangci-lint v2.14.0 reports 0 issues, and go run ./lint . reports no offenses.
• go test -race -shuffle=on ./pkg/runtime/ ./pkg/model/provider/anthropic/ ./pkg/modelsdev/ passes.
• An independent review found 0 critical and 1 important issue: the concurrent test missed a whole-request substitution. It's fixed here.

Deferred minor findings:

• The two requests are likely, not forced, to overlap.
• TestRoutedPricing creates a temp dir for every case.

[Copilot]

🟢 Approval recommended

The focused test-only changes cover the stated regressions, with no unresolved blocking issues.

No open issues 🎉

What changed in this PR

Adds test-only coverage for Claude 5.x lifecycle behavior as the first part of #4654, reusing existing infrastructure without production changes.

Changes:

• Pins Haiku 5.5 pricing at the 100,000-token boundary.
• Checks that routed serving-model identity and cost survive SQLite reload.
• Checks request isolation between concurrent sessions sharing one Anthropic client.

File	Description
pkg/​runtime/​routing_pricing_test.go	Verifies routed model and cost persistence.
pkg/​modelsdev/​snapshot_test.go	Checks Anthropic and Bedrock pricing boundaries.
pkg/​model/​provider/​anthropic/​features_integration_test.go	Checks shared-client session isolation.

_{Review effort: Balanced}