[leehack]

Summary

• Root cause of Diagnose browser Cache.put internal errors observed in Web QA #1675: Chromium keeps Cache Storage in memory for Incognito and other off-the-record profiles, and that in-memory store rejects any entry of 256 MiB or more with UnknownError: Failed to execute 'put' on 'Cache': Unexpected internal error. Web QA smokes open browser.new_context(), which is such a profile, and they load models larger than 256 MiB. Persistent profiles store the same bodies.
• The bridge already handles the rejection correctly: model_cache_state is store_failed, no entry is written, the model still loads, and there are no page errors. No runtime change is needed. This PR documents the limit in doc/webgpu_bridge.md and on the website WebGPU bridge page.

Production-readiness scope

• User-facing scope: docs explain why an Incognito or Playwright-context Web app logs Model was not cached and fetches a model again on every load.
• Supported platforms/paths: docs only (Web bridge model cache).
• Unsupported or intentionally unavailable paths: in Chromium off-the-record profiles, model caching is unavailable for entries of 256 MiB or more. The load still succeeds without the cache.
• Out of scope / follow-ups: None. LiteRT-LM Web model caching was not examined.

Completeness checklist

• Declared scope is fully implemented, or explicitly reduced above.
• Unsupported platform/option combinations fail loudly with actionable diagnostics or are clearly documented as unavailable.
• Public API docs, README/website docs, examples, support matrices, and changelog entries are updated where relevant. (Docs clarification only, no behavior change, so no changelog bullet.)
• Regression coverage covers the original issue plus key negative/version-skew paths where relevant. (Docs-only. The deterministic browser repro is below.)
• Public capability changes update the validation coverage registry with a pack or a reasoned, issue-linked exemption; inventory checks pass (or N/A with scope). N/A: no capability change.
• Security/privacy review completed: no secrets, bearer tokens, signed URLs, or raw secret-bearing paths leak through logs, cache keys, metadata, errors, or snapshots.
• Follow-up work that is useful but not required for this PR is tracked in GitHub Issues, or explicitly marked "None" above.

PR type guidance

• Docs-only PR: no runtime behavior changes. The two Markdown files were validated with the Markdown-selected root suites.

Test Plan

• dart run tool/prepare_workspace.dart
• dart format --output=none --set-exit-if-changed . N/A: no Dart changes.
• dart analyze N/A: no Dart changes.
• dart test -p vm -j 1 --exclude-tags local-only N/A: docs-only. Ran the tool/ci/markdown_test_suites.txt suites instead: 729 passed, 4 skipped.
• dart test -p chrome --exclude-tags local-only N/A: docs-only.
• coverage-lib matrix row when lib/ behavior changes. N/A: no lib/ change.
• Other targeted/local-only validation: the deterministic browser repro below. Docs site build not run: the website edit adds no links.

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
Cache.put size probe (ad hoc, not a matrix row)	fetch response then cache.put at 255/256/600/1024/1100 MiB, ephemeral vs persistent	Playwright Chromium 149.0.7827.55 and Chrome 157.0.8103.38, macOS	Ephemeral: 255 MiB stored; ≥ 256 MiB rejected UnknownError … Unexpected internal error. Persistent: all sizes stored and matched	llamadart-evidence/issue-1675-lane6/probe-fetch.jsonl
Bridge cache probe (ad hoc)	bridge v0.1.83 wasm32 loadModelFromUrl(useCache: true), worker and direct	Chromium 149, LFM3-Audio-2B Q4_K_M (1,099,626,016 B)	Ephemeral: load OK, store_failed, 0 entries, exact warning, no page errors. Persistent: stored, then hit	bridge-probe-v0183-lfm3.jsonl

Review Notes

• Independent review status: N/A: standard-risk docs-only change.
• CI status / head SHA: green at 275e8229d3030086a6a5db0b4f06a5b785c4a21a (9 pass, 20 skipped by path selection).

High-risk regression review

• Classification: standard
• Implementation task: N/A
• Independent blocking QA task: N/A
• Exact head / current base: N/A
• Production-branch deletion, bypass, or miswire proof: N/A
• Affected-family real-model/artifact evidence: N/A
• Explicit unavailable-family or other N/A evidence: N/A
• Known PR-caused P1 regressions: 0
• Unresolved review threads: 0

Fixes #1675
Refs #1845

[leehack]

Landed through the 0.19.2 batch integration PR #1870 (rebase merge, one commit per constituent; this PR's pinned head was audited/CI-gated there).

[github-actions]

Chat app preview removed for leehack/llamadart-chat-pr-1851.