Week 4 Day 8: Fork, Steer, and Select

[skyzh]

Why

Week 4 ends its first seven days with observable-outcome evaluation, but it does not yet reconnect agent continuations to the tokenizer and KV-cache work from Weeks 1–3. Day 8 teaches one bounded control-state fork without claiming that completed filesystem effects were rewound.

Change

• add a solution-free Day 8 starter and compact reference implementation for one saved dense tokenizer/KV prefix, two fresh steered continuations, visible reuse facts, Day 7 evaluation, and explicit passing selection;
• carry a nonblank operator denial reason back as one ordinary model-visible observation while retaining legacy boolean approvals and writing no receipt for a denied effect;
• copy the already-modified disposable workspace and byte-identical completed receipt history into isolated branch roots before later branch-local effects;
• publish the Day 8 chapter, navigation, overview, README roadmap, and reviewer day map.

The exact main-relative scope is 14 paths, 1,114 insertions, and 34 deletions. Base: e4c3a2281bfd9ba7649270e15a0daca7fcb152a3; head: e7a3c8fdfa5db3a9b0de5a0e18e33183f66923b0; tree: 9be9f663f6605d7245e31c243042e1defc934e2e.

Review

• 452 reference tests passed with 8 intentional skips; cumulative Week 4 plus sync 117/117 and focused Day 8 6/6;
• copied learner test was byte-identical and expected-red 6/6 before cleanup;
• Ruff check/format, diff check, and manual mdBook build passed;
• cached Qwen/Qwen3-0.6B-MLX-4bit walkthrough rebound the paused transcript through the real tokenizer/dense-KV checkpoint, then forked two distinct steered prompts from one 55-token, 28-layer prefix; both branches reported 55 reused and avoided-prefill tokens;
• Sentinel's bounded editorial contribution clarifies the required save_checkpoint(...) plus create_checkpoint(...) rebinding sequence before either real-model branch. No non-editorial byte changed in that successor.
• the first test-only successor retains both branches' actual cache handles and compares them with is not, eliminating Python address-reuse flakiness while preserving shared-prefix-array identity checks;
• the cumulative test-only successor binds exact steering in the visible message and outcome, binds the outcome to the generator's actual reuse report, and rejects a foreign valid model checkpoint. Independent mutations for constant steering, blank recorded steering, fake zero reuse, and foreign-checkpoint acceptance all fail the ordinary suite;
• hosted macOS run 31645509138, job 94277994497, is terminal green on the exact head.

This Draft does not add effect-state rewind, a session server/tree, concurrent steering, COW/radix serving, Day 9 externalization, publication automation, a release, or a deploy.

AI-Assisted: GPT-5.6 Sol + Forge

Reviewed-by: GPT-5.6 Terra + Oracle (consistency)
Reviewed-by: GPT-5.6 Sol + Sage (correctness)
Reviewed-by: GPT-5.6 Sol + Tuner (performance)
Reviewed-by: GPT-5.6 Terra + Scholar (learner)