fix(anolisa): tail logs --limit to recent matches

[zyw02]

Why

anolisa logs capped the central JSONL scan at the first N matches (default 50). Once the log grew, later operations disappeared from the default view. anolisa bug already kept the newest rows; logs should use the same tail window so operators see the failure that just happened.

What changed

• Query cap: CentralLog::query keeps a sliding window of the most recent filter matches instead of stopping at the first N. Results stay chronological (oldest of that window first).
• Lock window: query takes a shared flock only long enough to snapshot a stable file length, then scans that prefix without holding the lock so a default --limit tail does not block concurrent append for an O(file-size) deserialize. limited_query_releases_lock_before_scanning_history parks after unlock, completes append while the scan is still parked, then resumes; the snapshot must not include that concurrent write.
• CLI help: --limit is documented as the most recent N (still max 1000; 0 returns none).
• Public rustdoc: CentralLog::query matches LogFilter.limit (most recent N, file order). A doctest locks 3 records / limit=2 → last two, oldest-first.
• Tests: query_limit_applies_after_filter expects the last two warn-or-above records; query_limit_tails_matches_inside_since_window tails inside --since; CLI combined-filter limit=1 keeps the later error, not the earlier warn.

Related issue

fixes #2617

User / Agent impact

anolisa logs and anolisa logs --limit N now show the newest matching records. Scripts that depended on seeing the oldest N hits from a large log will see the newest N instead (still oldest-first inside the window). --since behavior is unchanged except that --limit tails inside that window.

Risk and compatibility

• Public CLI, API, configuration, or documented behavior changed
• Privileged or security-sensitive behavior changed
• Cross-component contract changed
• Migration or rollback guidance is needed

--limit / LogFilter.limit now means most-recent N rather than first-N. That is the intended operator contract for an append-only audit log; library callers of CentralLog::query with a limit see the same tail semantics.

Validation

cd src/anolisa
cargo fmt --all -- --check
cargo clippy --all-targets --locked -- -D warnings
cargo test --locked -p anolisa-core --lib central_log
cargo test --locked -p anolisa-core --doc
cargo test --locked -p anolisa-cli query_combined_filter_honors_zero_and_one_limits
cargo doc --workspace --no-deps

Focused: query_limit_applies_after_filter, query_limit_tails_matches_inside_since_window, limited_query_releases_lock_before_scanning_history (50/50 locally), query_combined_filter_honors_zero_and_one_limits. Full cargo test --locked as root still hits the existing --install-mode user adapter notice suite (not caused by this diff).

Documentation and rollback

CLI --limit help and LogFilter.limit / CentralLog::query rustdoc now state most-recent N. No user-guide change (guide does not claim first-N). Revert this commit series to restore first-N scanning.

Review follow-up

Current head: b74a25d96e55e055a1823140f5e132a71b282d99.

This head keeps the tail --limit contract, the short lock window, aligned query rustdoc + doctest, and a deterministic overlap test (post-unlock hold, then append, then resume scan). No 5 ms sleep or 250 ms append SLO.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 03ee051aa8

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".

[kongche-jbw]

Review baseline: 435951fc25da0ca48cab9bc78924c51179d7200a...03ee051aa85571c9a371f5cfa33576e0baf8173c

[P2] 避免在完整历史扫描期间持有日志锁

src/anolisa/crates/anolisa-core/src/central_log.rs:288 把原先的 break
改为 pop_front 后，任何带 limit 的查询（CLI 默认 50）都要反序列化整个无界
JSONL。查询全程持有共享 flock，而 append 需要独占锁；大日志上运行
anolisa logs 会让并发审计写入等待 O(file-size) 扫描结束。

Possible direction: 在锁内取得稳定文件长度后释放锁并只读该前缀，或从尾部按完整行
反向扫描；补充受控并发测试，证明长历史尾查询不会在整个扫描期间阻塞 append。

[P2] 同步更新 query 的公开契约

src/anolisa/crates/anolisa-core/src/central_log.rs:230 仍写着 limit 保留
“first N”，但同文件的 LogFilter.limit 已改为“most recent N”。库调用方会看到
相反的公开契约，并可能据此写出错误预期。

Possible direction: 更新方法 rustdoc，并用一个 3 条记录、limit=2 的 doctest 固化
返回后两条且保持文件顺序的契约。

[zyw02]

@kongche-jbw both P2s on 03ee051aa85571c9a371f5cfa33576e0baf8173c are addressed on 568ce67915cc47cb3697607639b8e8629bc764ee.

• Lock window: shared flock is held only to snapshot a stable file length; the tail scan reads that prefix unlocked. limited_query_releases_lock_before_scanning_history overlaps append with a long-history limit=1 query.
• Public contract: CentralLog::query rustdoc now says most-recent N (file order), with a 3-record / limit=2 doctest.

Validation: cargo fmt --all -- --check, cargo clippy --all-targets --locked -- -D warnings, cargo test --locked -p anolisa-core --lib central_log, cargo test --locked -p anolisa-core --doc, cargo doc --workspace --no-deps.

Please re-review this head.

[kongche-jbw]

Review baseline: 435951fc25da0ca48cab9bc78924c51179d7200a...568ce67915cc47cb3697607639b8e8629bc764ee

[P2] 移除锁回归测试的墙钟竞态

src/anolisa/crates/anolisa-core/src/central_log.rs:984

这里只等待 5 ms 就假定查询已完成快照并进入扫描；若查询线程尚未被调度，
append 会先写入，查询便把新记录纳入快照，最终记录断言会随机失败。
同时，250 ms 的墙钟上限会把繁忙 CI 上的调度停顿误判为锁等待。

Possible direction: 用测试钩子或屏障确认 query 已越过 unlock 并暂停扫描，
再断言 append 在释放扫描屏障前完成；移除固定 sleep 和性能阈值，保留同一路径的回归覆盖。

[zyw02]

@kongche-jbw the P2 on 568ce67915cc47cb3697607639b8e8629bc764ee (wall-clock race in the lock overlap test) is addressed on b74a25d96e55e055a1823140f5e132a71b282d99.

limited_query_releases_lock_before_scanning_history now parks query after the shared flock is dropped, completes append while the scan is still parked, then resumes. The snapshot must be op-snapshot-tail, not op-during-scan. The 5 ms sleep and 250 ms append SLO are gone.

Validation: cargo fmt --all -- --check, cargo clippy --all-targets --locked -- -D warnings, cargo test --locked -p anolisa-core --lib central_log, cargo test --locked -p anolisa-core --doc, cargo test --locked -p anolisa-cli query_combined_filter_honors_zero_and_one_limits, cargo doc --workspace --no-deps. The overlap test passed 50/50 locally.

Please re-review this head.

[kongche-jbw]

Review baseline: 435951fc25da0ca48cab9bc78924c51179d7200a...b74a25d96e55e055a1823140f5e132a71b282d99

未发现阻塞问题。

Remaining risks:

• 未运行全量 cargo test --locked；并发文件锁只在 Linux/aarch64 环境验证，未覆盖 macOS。
• PR 当前仅显示 CLA 检查，尚无仓库 CI 结果可供交叉核验。

Verification scope:

• 已复查完整 diff、相邻实现与测试，以及全部 conversation、review、inline thread 和 timeline。
• cargo fmt --all -- --check
• cargo test --locked -p anolisa-core --lib central_log
• cargo test --locked -p anolisa-core --doc
• cargo test --locked -p anolisa-cli query_combined_filter_honors_zero_and_one_limits
• cargo clippy --all-targets --locked -- -D warnings
• cargo doc --workspace --no-deps
• limited_query_releases_lock_before_scanning_history 重复运行 50/50 通过。

[ikunkun-sys]

Review baseline: d80cd127c0f64133ad9d8277e51792d3fa36b2b9...b74a25d96e55e055a1823140f5e132a71b282d99 (synthetic merge 24c5419d6a8ccf9eca8592c45e0dd18ed641e96d)

[P2] 按仓库规则整理 fix 归因与提交历史

AGENTS.md:382-387

03ee051a 修复了 main 上 7a6e0d72 引入的 first-N 行为，但其 trailers 中没有必需的 Fixes:。随后 568ce679 修复 03ee051a 引入的锁窗口/文档问题，b74a25d9 又修复 568ce679 引入的测试竞态；这两个 review follow-up 仍是独立的 fix commits。

这与仓库规则冲突：main 上已有缺陷需要正确的 Fixes: 归因，同一 PR 内引入的问题必须 fixup/autosquash，不能保留独立 fix commit。当前最终代码行为验证通过，但合并前仍需补齐归因并整理为原子提交历史。

Verification: Rust 1.88.0、Linux x86_64 的 synthetic merge 上，workspace fmt、Clippy、docs、cargo test --workspace --locked 均通过；锁重叠测试重复 50/50 通过。

[ikunkun-sys]

Review baseline: b74a25d96e55e055a1823140f5e132a71b282d99 (unchanged since review #4956800515)

Requesting changes for the verified blocking item from that review:

[P2] 按仓库规则整理 fix 归因与提交历史

03ee051a 修复 main 上 7a6e0d72 引入的缺陷，但没有 Fixes:；568ce679 和 b74a25d9 又作为独立 fix commits 修复本 PR 前序提交引入的问题。该历史仍违反 AGENTS.md:382-387 的强制归因与 autosquash 规则。

当前代码验证和远端 CI 均已通过；本次 request changes 只阻塞上述提交历史问题。