[leehack]

Summary

• litert-lm-chat-features-smoke reads the bundle's embedded chat template and records the thinking row as not_applicable (with reason) when the template never reads enable_thinking and closes the think block in its generation prompt, instead of failing. Other templates keep the assertion.
• Corrects model-families.md: Qwen3.5 .litertlm renders the bundle's own template (#1134), ModelParams.chatTemplate is ignored, and the _int8, -VL_int8 and _int8_gpu_composite_v0190 bundles produce no thinking (_dynamic_wi8_afp32 needs LiteRT-LM 0.20+).

Refs #1827; left open for the design question (applying a thinking-capable template to Qwen3.5 text turns).

Production-readiness scope

• User-facing scope: Docs correction; local-only smoke reports the thinking row accurately.
• Supported platforms/paths: Tooling and docs only; no lib/ change.
• Unsupported or intentionally unavailable paths: Qwen3.5 .litertlm thinking stays unsupported, now documented and reported as N/A with the reason.
• Out of scope / follow-ups: Library template choice or typed rejection: #1827.

Completeness checklist

• Declared scope is fully implemented.
• Unsupported option documented (model-families.md).
• Docs updated; changelog N/A (docs correction, no user-visible runtime change).
• Unit tests cover templates that honour enable_thinking, the always-closed template, and unreadable/other templates.
• Validation registry: N/A (no capability change).
• No secrets involved.
• Follow-ups tracked in LiteRT-LM Qwen3.5 bundles never emit thinking (embedded template pre-closes <think>), failing litert-lm-chat-features-smoke on CPU and GPU #1827.

PR type guidance

• Bugfix PR: root cause in LiteRT-LM Qwen3.5 bundles never emit thinking (embedded template pre-closes <think>), failing litert-lm-chat-features-smoke on CPU and GPU #1827; regression test is the template-decision unit test; real smoke below.

Test Plan

• dart run tool/prepare_workspace.dart
• dart format --output=none --set-exit-if-changed .
• dart analyze
• dart test -p vm test/unit/tooling/ (only the 6 windows_registry_qualification_test.dart failures, pwsh missing on this Mac)
• Full VM/Chrome suites: N/A, no lib/ change; left to CI
• coverage-lib: N/A, no lib/ change
• Real smoke, litert-lm-chat-features-smoke, Qwen3.5-0.8B-VL int8, CPU, macOS arm64: exit 0; thinking row status: not_applicable with reason, all other checks pass.

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
litert-lm-chat-features-smoke	chat, cutoff, tool loop, history, required-tool rejection; thinking N/A	macOS arm64, Qwen3.5-0.8B-VL_int8 .litertlm, CPU	PASS	RESULT litert_lm_chat_features with thinking not_applicable
high-risk-exact-head-independent-qa	exact-head audit of eda933e vs ec4bf4f	independent sessions	PASS	see block below; evaluator exit 2 (unverifiedPrerequisites)

Review Notes

• Independent review status: accepted (audits above); draft retained, not marked ready
• CI status / head SHA: eda933e (origin/main ec4bf4f merged); CI green; exact head 2d7b3a1 audit passed

High-risk regression review

• Classification: high-risk (surface: regressionPolicy, via tool/testing/test_matrix.dart and doc/testing_matrix.md)
• Implementation task: this PR
• Independent blocking QA task: fresh independent audit sessions: full audits passed at e159c34 and 2d7b3a1 (no P1/P2, mutations caught); exact-head re-audit passed at eda933e (clean merges, PR diff patch-id equals the audited 2d7b3a1 diff, CI 34 pass/9 skipped)
• Exact head / current base: eda933e / ec4bf4f
• Production-branch deletion, bypass, or miswire proof: thinkingSupportForTemplate mutation (closed-think check replaced by a constant) fails reports a template that always closes the think block as N/A; reverting the covers text fails documents the LiteRT-LM chat smoke thinking N/A
• Affected-family real-model/artifact evidence: real Qwen3.5-0.8B-VL int8 CPU smoke: exit 0, thinking row not_applicable with reason; the audits confirmed Qwen3-0.6B and Gemma 4 E2B still assert thinking
• Explicit unavailable-family or other N/A evidence: N/A (no structured-output surface)
• Known PR-caused P1 regressions: 0
• Unresolved review threads: 0

Readiness evaluator (tool/testing/high_risk_readiness.dart, evidence with schema_version 2.0.0): unverifiedPrerequisites, exit 2, as the contract expects locally.

[leehack]

Landed via batch integration PR #1843 (rebase merge), one commit per constituent.