[Docs] Require artifact hygiene and full review for large code changes

[hsliuustc0106]

Why

Port the review safeguards from merged ThinkFlowLab/system1-omni#105 to System1-Agents: redundant generated output should not obscure a change, and large authored-code diffs need a complete contributor review.

How

• CONTRIBUTING.md defines the >3,000 changed authored-code line threshold: additions + deletions from the merge base in source, tests, and build/validation scripts. Documentation, generated output, lockfiles and static fixtures stay out of that count but remain in review scope.
• Require full self-review, a split rationale/component map/review order, and validation for each component and changed interface. Keep large PRs draft until contributor self-review is complete. Size alone is neither a defect nor a requirement for GPU/expensive runs.
• The existing self-review skill and contributor guide apply artifact hygiene to every review, including quick prechecks. Preserve maintained fixtures/replay inputs and durable raw evidence; check actual consumers, hashes, links and replay/documentation commands when cleaning up output.
• Reconcile the existing demo-upload wording with the artifact policy. No separate precheck or reviewer skill exists in this repository.

What

Two Markdown files change, plus the isolated three-file CI repair below. Authored-code count: 4 additions + 0 deletions. Total diff: 66 additions + 2 deletions = 68 changed lines. Base/merge base: 621681b5d59a779e3e93b9524c1772a1970fe7d7. Head: 945dc4e8e637fbfc5ae6eedbd8f19d55c717ac2e.

Isolated CI repair

The original documentation head inherited three failures already present on exact main 621681b: ServedLayaModel and the ServedStub test double omitted bills_input_tokens, causing billing-contract failures and an AttributeError in the tool wrapper. The earlier main CI run and original PR run show the same failures.

A separate follow-up commit reuses only the focused correction from #37 at 13d9613d896c9ee8a8f663614e5c334d2ccdab09: explicitly declare both served classes non-billable at Jev rates and add the real served backend to the billing-contract expected map. Existing contract and tool-provenance tests cover the regression. No Laya dependency, precision or other inference changes from #37 are included.

PR #45's video-evidence edits are left intact and separate. Its overlapping CONTRIBUTING/self-review patch applies cleanly on top of this change; no video requirement is weakened.

Verification

• Passed git diff --check against exact source contents.
• Passed local Python assertions for added relative links and heading anchors, balanced Markdown fences, and policy coverage (counting exclusions, draft/full-review gate, split/review map, evidence/fixture preservation, consumer/hash/replay checks).
• Passed git apply --check --include=CONTRIBUTING.md --include=.agents/skills/self-review/SKILL.md ../agents-pr45.patch for the pending [Docs] Require application, agent and Omni video demos #45 overlap.
• Reviewed the complete two-file diff against the authoritative #105 diff; no actionable findings.
• Verified the original published remote comparison contained exactly the two intended Markdown files. The subsequent scoped repair adds four Python lines across three files.
• Follow-up repair: all three changed Python files parse; dependency-free CPU assertions executed their actual class declarations and reproduced missing billing attributes before the fix and explicit False after it; the served expectation is present in the regression contract. These focused checks are not a full pytest run.
• Not run locally: Ruff, ty, pytest, smoke, browser/system tests or model/GPU runs. The partial cloud checkout lacks the project dependencies. No local full-suite pass is claimed.
• Observed exact-head CI run on 945dc4e8e637fbfc5ae6eedbd8f19d55c717ac2e: all three core jobs (Linux 3.11/3.13, Windows 3.11) passed with 606 tests passed and 43 skipped each; full passed with 615 tests passed and 33 skipped. Smoke passed in all four jobs, along with configured lint/type/build checks. Browser is intentionally skipped for pull-request events. No failed checks remain in this run.

Demo / evidence

No generated run artifacts or media are added or removed. The policy portion needs no agent demo. The four-line CI repair restores an existing served-model billing contract and is validated by automated regression tests; no live inference/performance claim or model/GPU run is made.