Harden model streaming for 2048 demo

[willamhou]

Summary

• raise streamed tool-call max tokens to 4096 with bounded env override
• raise streaming timeout defaults and expose demo-specific overrides
• make the 2048 recorder reproducible with an isolated HOME, temporary demo skill, retry cleanup, and stronger post-run validation

Verification

• cargo fmt --check
• git diff --check
• node scripts/check-secrets.js
• bash -n docs/demo/record-2048-demo.sh
• docs/demo/record-2048-demo.sh --dry-run
• docs/demo/record-2048-demo.sh --redaction-self-test
• cargo test --lib
• cargo build --bin deepseek
• live: docs/demo/record-2048-demo.sh

[github-actions]

DeepSeekCode review of PR #34 (Harden model streaming for 2048 demo)

Let me examine the repository structure and the changed files to understand the PR.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8f74447e99

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Reset the demo repo fully between retry attempts

When a model attempt fails, reset_demo_attempt only removes index.html/styles.css/app.js and runs git reset --hard HEAD, which leaves any other untracked files from that attempt in place. Because write_file can still create additional paths, a retry may run against a non-empty workspace and validation can pass with stale artifacts, undermining the script’s reproducibility guarantee for “empty repo” demo captures.

Useful? React with 👍 / 👎.