test(lark): pin the manager context window's bounds and order

[hahahahahayesyeseys]

Goal And Delivered Outcome

• Goal/source and gap: Lark manager: decouple context capture from Turn authority #4318's capture/authority split is implemented and its read side is manager_context_materials(). Its contract is three named limits -- at most MANAGER_CONTEXT_ITEM_LIMIT messages, at most MANAGER_CONTEXT_CHARACTER_LIMIT characters, consumed newest-first so the surviving set is returned in chronological order. None of the three was tested: searching the suite for the constants, the per-item cap or context_window_exceeded returns nothing, and only the _for_ids replay sibling is exercised directly. So the issue's "bounded" clause had no enforcement behind it.
• Observable before -> after: before, three mutations of the window all passed unnoticed or were caught only indirectly; after, each is caught by a named case (see the mutation row). No production behavior changes.
• Issue reference: Refs Lark manager: decouple context capture from Turn authority #4318. Tests only.

Scope And Continuation

• Delivered: three cases in the file that already owns manager context and its retention receipts, written from the contract rather than from recorded output -- window keeps the newest and restores arrival order, the character budget is spent from the newest side so the oldest survivor is the clipped one, and addressed/current messages never enter context while replayed history does.
• Remaining work, named rather than folded in: MANAGER_CONTEXT_RETENTION_LIMIT = 32 and MANAGER_CONTEXT_MAX_AGE are partially covered by the retention test in the same file, and the context_window_exceeded label still has no case of its own. Those belong to the retention path, not to this window, and none of them is a blocker here.
• Boundary / successor: this adds coverage to a merged contract (fix(lark): preserve context without granting Turn authority #4319); no follow-up PR is implied.

Validation

• Tested revision: 4496de4 (current head).
• Run state: finished. Input classes: synthetic.

Check kind	Result	Public-safe evidence / limitation
unit	passed	tests/extensions/test_lark_manager_reply_delivery.py: 8 passed, of which the 3 added here are new
unit	passed	tests/extensions: 952 passed at this head -- the added file is collected with its neighbors unchanged
regression_parity	passed	mutation, not assertion count: take the oldest slice instead of the newest -> only the window case fails; drop the final chronological restore -> all three fail; let addressed messages into the window -> only the exclusion case fails. Each case therefore binds a distinct property
static	passed	ruff check clean on the changed file; py_compile clean; git diff --check clean; loopx check --scan-path on the changed file reports 0 errors

• Base note, stated rather than left to inspection: this branch is rooted at 04ba65ac1, which is behind main. The two files it depends on -- loopx/extensions/lark/manager_context.py and this test file -- are byte-identical between that revision and current main (verified by blob hash), so nothing here is measured against stale behavior. It is based there because pushing a current-main branch from this fork requires a token scope this author's token does not carry, while a fork sync is pending.

Type of Change

• Bug fix
• New feature
• Breaking change
• Refactoring (no functional changes)
• Documentation update
• Test update

LoopX Area

• Control plane (goals, todos, quota, scheduler, registry, runtime)
• Benchmark boundary (adapters, runners, verifiers, scoring, evidence)
• Capability or extension (providers, adapters, skills)
• Public docs or presentation surface (README, protocols, dashboard)
• Build, packaging, installer, or CI
• Host or runtime integration

Technical Direction

• Direction / acceptance reference: Lark manager: decouple context capture from Turn authority #4318 acceptance clause 3 ("a later authorized manager Turn may read a bounded recent context window") -- this is the test that makes "bounded" a checked property instead of a constant.

Shared-authority RFC fixture impact

• Production-scale fixture schema: N/A -- no persisted schema, receipt or cursor contract changes.
• Semantic dimensions changed, or reviewed no-impact rationale: N/A -- tests only.
• Provider conformance arms run: none; no provider or AuthorityStore selection changes.
• Read-only legacy/file/PostgreSQL three-arm rehearsal: not applicable.

Boundary Checklist

• Neither the diff nor this PR body discloses private state, credentials, raw traces, internal links or local machine paths.
• I did not duplicate maintainer-owned work.
• I kept the change scoped to the linked issue.
• I completed the visual evidence section for UI changes, or marked UI impact none. -- UI impact: none; no presentation surface is touched, the change is a test file.
• Every commit includes a DCO Signed-off-by trailer.

[huangruiteng]

动机

#4318 已交付 Lark manager 的“context capture 与 Turn authority 分离”，但它要求后续 authorized Turn 只读取有界的近期上下文。现有测试覆盖 retention、receipt-bound replay 和 capture/authority 分离，却没有直接调用 manager_context_materials() 来约束三个关键性质：最多 8 条、总计最多 4000 字符、从最新消息一侧消费预算后按时间顺序返回。这个 PR 只补这一真实 coverage gap，不改变生产行为。

改动思路

三个测试都从公开的持久行为设计 oracle，而不是照抄当前输出：超出 item limit 时必须丢最旧三条并保持 survivor arrival order；字符预算必须优先保护最新消息，只有最旧 survivor 可被截断；current/addressed authority-bearing 消息不得同时作为 quiet context，而 historical_context_only replay 仍可进入窗口。测试直接走生产 read helper，没有新增 fixture framework、mock owner 或第二套规则。

具体改动

• 在既有 test_lark_manager_reply_delivery.py 中增加一个小型 _chat_window synthetic fixture。
• 增加 newest-window/order 用例，直接绑定 MANAGER_CONTEXT_ITEM_LIMIT。
• 增加 newest-side character-budget 用例，绑定 MANAGER_CONTEXT_CHARACTER_LIMIT 和截断方向。
• 增加 current/addressed exclusion 与 historical replay inclusion 用例，守住 context/authority 分界。

关键测试讲解

test_context_window_keeps_the_newest_messages_in_arrival_order 同时防止“取 oldest slice”和“返回 reverse order”；test_context_character_budget_is_spent_from_the_newest_side 让 8×1200 字符触发 4000 限制，期望 chronological sizes 为 [400, 1200, 1200, 1200]；test_context_materials_exclude_the_addressed_and_the_current_message 用同一 projection 同时证明 current/addressed 被排除而 explicit historical context 被保留。

对主干的风险

这是 test-only diff，无 schema、runtime、permission、quota 或 UI 变化。独立覆盖扫描确认 merge base 上只有 _for_ids replay sibling 被直接调用，item/character constants 没有测试；同作者当前只有这一项 open PR，没有 same-shape batch farming。exact head 的 changed file suite 8 passed，Ruff、py_compile、diff check 与 public boundary scan通过；标准 canary通过。由于该分支基于旧 merge base，我还把 patch 干净合入最新 origin/main@0d464d2ac，保留主干后来新增的 proposal-delivery test，合并态 suite 6 passed。

有一个非阻塞的证据表述需要更正：PR body 说 production file 和 test file 在旧 base 与当前 main 间都 byte-identical；独立 blob hash 显示 production manager_context.py 的确相同，但 test file 不同（主干后来新增了其它测试）。这不影响 patch 的可合并性或测试结论，因为合并态已单独验证，但建议作者把 body 改成“production file identical，test patch cleanly integrates”。

我的整体评价

这 82 行不是一次性 walkthrough：它给已发布的 manager context window 增加三条互补、可产生突变敏感性的长期 contract tests，位置与现有 retention/replay suite一致，且没有重复覆盖或批量灌水。除 PR body 的非阻塞 base 描述修正外，未发现 blocker，我赞成在当前 exact head 上 APPROVE；合并动作仍由 maintainer 决定。

English verdict: APPROVE - head 4496de4; three durable tests now pin newest-first bounds, chronological output, and context/authority exclusion, while the stale-base body note should be corrected; exact-head 8/8 and latest-main integration 6/6 passed with static and boundary checks clean.