[ranxianglei]

Closes #4194

What

4d50f9c2 (#4184, implementing #3900 "reserve output headroom on every wire") dropped the Anthropic exemption, so every wire's effective window now carries min(max_tokens, cap × window). It realigned anthropic-beta-window / context-overflow / preflight-max-output-tokens but left five other suites asserting the pre-#3900 numbers. This PR realigns those suites — no source changes.

Per-suite mechanism (verified by local repro on fe5ba59, then fixed)

Suite	Failure	Mechanism
tests/codex-window-align.test.ts	0/8 — every limit exactly −1024 (= fixture max_tokens)	The clamped/declared window surfaces carry the per-request reservation. All assertions now <base> − FIXTURE_MAX_TOKENS (new named constant).
tests/side-request-isolation.test.ts	1 — anthropic row expected blocked=false under tolerance	The shared body carried max_tokens, so anthropic now reserves and the boundary flips. Pure guard-math rows now run without an output-budget field; added an explicit anthropic reservation boundary pair mirroring the OpenAI case (both directions, #3617 gate discipline).
tests/external-summary-autofold.test.ts	3 — nudge never fired	Proxy-direct requests miss the route table → window resolves as fallback → #3900 reservation runs → the pre-existing #282 fallback floor (handle.ts: nativeFromFallback && reserved < 100_000) re-inflates the post-reservation window (48976 → 100000) → usage judged at 42% instead of 85.8% → pressure band cold. Fix: give the fixture model an authoritative registry window (setRegistryForTest) so the window is non-fallback and no floor applies.
tests/fork-nudge-deadzone.test.ts	1 — lane proof	Not window math: the new [headroom] … nudge/preflight judge … log line contains the word "nudge" and logs before the diagNudge decision line, so the selector picked it up first. Selector now pins the [<id>] nudge prefix. (Probe dump confirmed the decision line itself reads nudge idle: usage=32% — same as pre-#3900.)
tests/plugin-fork.test.ts	1 — 398976 !== 400000	Status contextLimit is session.metadata.effectiveContextLimit (src/plugin.ts:1611) — the window nudge/preflight judge against — which carries the reservation by design. Assertion now 400000 − 1024.

Behavior-change disclosure (from #3900 itself, surfaced here for the record)

On the Anthropic wire, small fallback-sourced windows (< 100K after reservation) are now effectively floored to 100K by the pre-existing #282 rule — e.g. a 50K guessed window judges usage against 100K instead of 48976. This is a consequence of #3900 × #282, ships with master already, and is out of scope for this test-only PR — flagged per the behavior-drift rule. If the owner wants that interaction changed, it belongs in a follow-up against #3900.

Verification

• Local targeted runs on this branch: 101/101 across all five suites (codex-window-align 8, side-request-isolation 33, external-summary-autofold 8, fork-nudge-deadzone 3, plugin-fork 49).
• npm run typecheck clean.
• Full-suite verdict = remote CI test lanes on this head.

[github-actions]

📦 Built Package Artifact

Branch: 2026-10-22_realign-post3900-window-expectations (bc23a5e)

Option A — Install this exact build (recommended)

npm install -g [email protected]

This exact version is immutable and reproducible. The rolling pr tag tracks the newest test build across all PRs.

Option B — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf billion-context-pr4208.tgz
npm install -g package

---

Install instructions are refreshed on each push.

[ranxianglei]

Closing as superseded: while this PR was in flight, #4198 merged (2026-10-21T16:37Z) reverting #4184/4d50f9c2 entirely — master is back to the pre-#3900 reserve-on-every-wire semantics, and v0.1.204 shipped from that state. This PR realigned five suites to the POST-#3900 numbers, which is now the wrong direction: the CI failure on this head (codex-window-align actual=272000 vs expected 270976 etc.) is the merge of these expectations with the reverted code, not a bug in either side.

Per #4198's stated plan, #3900 stays open for a redo with different semantics (record the un-reserved value in effectiveContextLimit, apply the reservation at the band layer, update all affected tests + goldens in one pass). If that redo lands, the per-suite mechanisms worked out here are still reusable reference:

• codex-window-align / plugin-fork status: which surfaces carry the reservation and by how much (base − max_tokens under cap 0.25);
• external-summary-autofold: the [BUG] acp-proxy upstream HTTP 400: Request with 1,198,269 tokens exceeds the model's 1,048,576 token limit, consistently | [BUG] acp-proxy upstream HTTP 400：请求 1,198,269 tokens 超过模型上限 1,048,576，一直复现——是插件压缩未生效，还是 DSH 的问题？ #3900-reservation × bili出现连续压缩情况 #282-fallback-floor interaction (post-reservation fallback windows <100K get floored back up, keeping nudge pressure cold) — worth deciding explicitly in the redo;
• fork-nudge-deadzone: the [headroom] diagnostic line contains the word "nudge" and shadows a bare-substring log selector — any redo that keeps that line needs the [<id>] nudge prefix pin;
• side-request-isolation: the guard-math rows vs reservation-boundary rows split (both directions per Test Blind Spot Retrospective: Reinforcing Negative Assertion Discipline, Tightening Fake Upstream Strictness, Thin Inst | 测试盲区复盘:补反向断言纪律、假上游严格化、薄安装冒烟、形态矩阵与生命周期覆盖 #3617).

No source changes were involved; the branch 2026-10-22_realign-post3900-window-expectations remains available for reference.