fix: prevent Qwen 2.5 WASM out-of-memory loads

[konard]

Summary

Fixes #17.

Qwen2.5 1.5B was advertised as fitting CPU-only Safari because the model selector estimated ONNX runtime memory from download bytes but omitted ONNX Runtime's additional WASM heap/session-construction overhead. Loading the 1.22 GB q4f16 graph then terminated inside Emscripten with the opaque Aborted() shown in the issue.

This PR:

• applies backend-aware peak-memory estimation: the existing model/activation estimate is retained for WebGPU, while WASM includes conservative graph/allocator/session heap headroom;
• classifies Qwen2.5 1.5B as too large on a 2 GiB CPU/WASM budget, while keeping q4f16 available on capable WebGPU devices;
• permits WebGPU → WASM retry only when the exact model/dtype also passes the CPU fit calculation;
• chooses the smallest fallback artifact by measured bytes (Qwen's q4 file is larger than q8, so dtype labels are not a safe size ordering);
• archives the original issue screenshot and provides a complete requirement matrix, timeline, root-cause analysis, upstream research, alternatives, and verification in docs/case-studies/issue-17/;
• adds the required changelog fragment.

Reproduction

Before this fix:

1. Open the deployed app in Safari (WASM/CPU backend).
2. Select Qwen2.5 1.5B Instruct at q4f16.
3. Observe that the card says the model fits and begins the ~1.1 GiB download.
4. Session construction ends with Error: Aborted(). Build with -sASSERTIONS for more info.

Original evidence:

Automated regression

web/src/models/models.test.ts now evaluates the real Qwen2.5 1.5B catalog entry with a CPU-only 2 GiB budget and asserts it is too-large; the old code returned tight and allowed the crashing load. A companion assertion verifies q4f16 remains selected with a usable WebGPU adapter.

Verification

Passing locally:

• npm test --prefix web -- --run — 39 tests
• ./scripts/build-wasm.sh
• npm run build --prefix web — TypeScript + production Vite build
• cargo fmt --check
• cargo test — 19 unit/integration tests + 3 doctests
• cargo clippy --all-targets --all-features
• node scripts/check-file-size.mjs
• git diff --check

Known repository-wide tooling issue: npm run lint --prefix web cannot start because ESLint 9 is installed but the repository has no eslint.config.js/mjs/cjs. This predates and is unrelated to the changed files; TypeScript compilation and all configured CI checks remain covered above.

Case study

See docs/case-studies/issue-17/ for preserved evidence and the full analysis.

[konard]

Working session summary

Implemented and finalized PR #18.

Key changes:

• Corrected WASM peak-memory estimation so Qwen2.5 1.5B is not offered on unsafe CPU-only devices.
• Guarded WebGPU-to-WASM fallback using the same memory-fit calculation.
• Selects fallback variants by actual artifact size.
• Added regression tests for CPU and WebGPU behavior.
• Archived the original screenshot and added the full issue case study.
• Added a changelog fragment and removed the WIP placeholder.

Verification:

• 39 web tests passed.
• Production WASM and Vite builds passed.
• Rust tests, doctests, formatting, Clippy, and file-size checks passed.
• All fresh GitHub CI, deployment build, and browser E2E workflows passed on commit 5f2a2c9.
• Worktree is clean; PR is mergeable and ready for review.

---

This summary was automatically extracted from the AI working session output.

[konard]

🤖 Solution Draft Log

This log file contains the complete execution trace of the AI solution draft process.

💰 Cost estimation:

• Model: GPT-5.6 Sol
• Provider: OpenAI
• Public pricing estimate: $14.741969

📊 Context and tokens usage:

• 281.7K / 200K (141%) input tokens, 22.6K / 128K (18%) output tokens

Total: (281.7K + 10.9M cached) input tokens, 22.6K output tokens, $14.741969 cost

🤖 Models used:

• Tool: OpenAI Codex
• Requested: gpt-5.6-sol
• Thinking level: medium (~15999 tokens)
• Model: GPT-5.6 Sol (gpt-5.6-sol)

📎 Log file uploaded as Gist (3032KB)

• View complete solution draft log

---

Now working session is ended, feel free to review and add any feedback on the solution draft.

[konard]

✅ Ready to merge

This pull request is now ready to be merged:

• All CI checks have passed
• No merge conflicts
• No pending changes

---

Monitored by hive-mind with --auto-restart-until-mergeable flag