[leehack]

Summary

• LiteRT-LM Web getContextSize now returns the executor limit the load probe found (_reportedContextTokens) instead of the requested ModelParams.contextSize, which a fixed KV cache clamps (Qwen3 0.6B on WASM CPU: requested 65536, enforced 4096). Falls back to the requested size when no probe result exists.
• Report LiteRT-LM Web token-limit truncation and run automatic tool loops there #1891 re-audit item in #1861. Refs 0.19.x audit follow-ups (index) #1861 (tracker not ticked).

Production-readiness scope

• User-facing scope: engine.getContextSize() on LiteRT-LM Web withTextHistory() loads reports the enforced limit.
• Supported platforms/paths: LiteRT-LM Web. Native and other backends untouched.
• Unsupported or intentionally unavailable paths: runtimes without the probe keep returning the requested size.
• Out of scope / follow-ups: None.

Completeness checklist

• Declared scope is fully implemented.
• Unsupported platform/option combinations: unchanged fallback.
• Changelog (CHANGELOG.md and website recent-releases) updated; no skills or support-matrix claim changes.
• Regression coverage: new browser test, failing before the fix (expected 4096, got 65536), plus the unprobed fallback case.
• Validation coverage registry: N/A, no capability change.
• Security/privacy review: no secrets involved.
• Follow-ups: None.

PR type guidance

• Bugfix PR: root cause is getContextSize reading the requested ModelParams.contextSize while the executor enforces its own clamped copy; regression test above; affected platform is LiteRT-LM Web.

Test Plan

• dart format --output=none --set-exit-if-changed .
• dart test -p chrome test/unit/backends/litert_lm/litert_lm_backend_web_test.dart (172 passed)
• coverage-lib: N/A, the file is Web-only and not in VM coverage

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
static-format-analyze	format	macOS	PASS
root-chrome	LiteRT-LM Web backend suite	Chrome, fake @litert-lm/core engine	PASS
high-risk-exact-head-independent-qa			N/A	pending, draft

Review Notes

• Independent review status: pending; draft until audited
• CI status / head SHA: see checks

High-risk regression review

• Classification: high-risk (backendRuntime)
• Implementation task: lane 1861b (this branch)
• Independent blocking QA task: pending
• Exact head / current base: pending
• Production-branch deletion, bypass, or miswire proof: reverting the _reportedContextTokens ?? term fails the new test (4096 expected, 65536 returned)
• Affected-family real-model/artifact evidence: N/A, no Web hardware run; fake engine only. The real-model litert-lm-web-truncation scenario was not re-run.
• Explicit unavailable-family or other N/A evidence: real LiteRT-LM Web model run unavailable here
• Known PR-caused P1 regressions: 0
• Unresolved review threads: 0

[leehack]

Landed via batch integration PR #1912 (rebase merge, main 2cfb936). The integration audit passed at a9c7772, and this PR's audited head is recorded in #1912's constituent table.

[github-actions]

Chat app preview removed for leehack/llamadart-chat-pr-1901.