fix(tokenless): fail compare on missing sessions

[zyw02]

Why

tokenless stats summary --compare treated unknown session IDs as a successful 0% comparison. A mistyped --session-id or a dual-run that never recorded stats looked like "no savings" instead of "this session does not exist." stats diff --session and the Python TokenlessStats.compare client already fail closed on the same case.

What changed

• --compare empty-session contract: if either named session has no records, the CLI now exits 1 with No records found for … and does not print a 0% text or JSON report.
• --limit 0: stats summary --limit now uses parse_positive_usize like stats diff, so --limit 0 fails at parse time instead of looking like a missing compare session.
• Docs: component README (en/zh) and user-guide CLI/measurement pages now summarize both public contracts: missing --compare sessions fail closed, and stats summary --limit must be a positive integer.
• Adjacent coverage: both sides missing, only baseline missing, only tokenless missing, both populated (text + --json), control-character session IDs (debug-escaped like stats diff --session), and --limit 0 on populated --compare.
• Binary proof: tokenless-cli integration tests run the same --compare invocations against an isolated TOKENLESS_STATS_DB.

Related issue

fixes #2673

User / Agent impact

tokenless stats summary --compare <baseline> <active> now fails when a session was never recorded, matching tokenless stats diff --session and TokenlessStats.compare. Dual-runs that actually wrote stats are unchanged. stats summary --limit 0 is now rejected at parse time, matching stats diff --limit 0.

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

CLI --compare on an empty session now errors instead of exiting 0 with a zeroed report. That is the bugfix. Callers with two recorded sessions keep the same comparison output. Python compare already raised StatsNotFoundError for this case. --limit 0 now fails at clap parse instead of querying.

Validation

Current head: 6278f1db9cfcff64bb8cae83ff81009e09664410

DIRTY rebase onto upstream main a4cb23ea27c9645d7294b7afe65309262e6a0e4b from previous head 72e92ed2ad094147de313ed8820b7bdcc3a4d05f. One conflict file: src/tokenless/crates/tokenless-cli/src/tests/main_tests.rs. Resolution kept upstream persist snapshot / enable-disable tests (stats_persist_snapshot_*, run_command_stats_{enable,disable}_does_not_persist_env_overrides) and this PR's missing_compare_sessions_* tests plus seed_compare_record. main.rs auto-merged: upstream persist_stats_enabled / load_from_file path kept, with this PR's fail-closed --compare and parse_positive_usize on --limit replayed on top.

The previous 3-commit history was squashed into one atomic commit on this rebase, per AGENTS.md §13 and review.

cd src/tokenless
cargo fmt --all --check
cargo clippy --workspace --locked -- -D warnings
cargo test --workspace --locked -- --test-threads=1

Rebased head 6278f1db9cfcff64bb8cae83ff81009e09664410: fmt/clippy clean; workspace tests 606 passed, 2 ignored. Focused tokenless-cli: 9 unit tests (compare fail-closed, persist snapshot, --limit 0) + 2 integration tests (stats_summary_compare_rejects_missing_sessions, stats_summary_compare_rejects_zero_limit) passed.

Exact-head proof (TOKENLESS_STATS_DB pointed at an isolated file):

tokenless stats summary --compare missing-a missing-b
# Error: No records found for baseline session "missing-a" and tokenless session "missing-b"
# exit=1

tokenless stats summary --limit 0 --compare missing-a missing-b
# invalid value '0' for '--limit <LIMIT>': value must be greater than zero
# exit=2

Review follow-up (--limit 0 on populated compare sessions): clap rejects with value must be greater than zero and does not emit No records found. Covered by stats_summary_cli_rejects_zero_limit and stats_summary_compare_rejects_zero_limit.

Review follow-up (docs): README and user-guide now document both public CLI contracts requested on 1ceb816.

Documentation and rollback

User-guide notes in docs/user-guide/{en,zh}/token-saving/tokenless/measuring-savings.md and cli-reference.md, plus src/tokenless/README.md / README_zh.md, state that a missing session fails instead of reporting 0%, and that stats summary --limit must be a positive integer. Revert the commits on this branch to restore the previous zeroed success report, unvalidated --limit 0, and prior docs.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 45d5fe9e66

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".

[Forrest-ly]

Code Review — commit 45d5fe9

Overall

The core fix is correct and well-tested. stats summary --compare now fails closed on missing sessions, aligned with stats diff --session (same No records found for … message style, exit code 1, plain-text stderr error even in --json mode) and with the Python TokenlessStats.compare (StatsNotFoundError). Verified locally at head 45d5fe9:

• cargo fmt --all --check ✅
• cargo clippy --workspace --locked -- -D warnings ✅ (0 errors)
• cargo test --workspace --locked -- --test-threads=1 ✅ (588 passed, 2 ignored)
• Manual binary check: stats summary --compare missing-a missing-b → Error: No records found for baseline session "missing-a" and tokenless session "missing-b", exit=1, no 0% report on stdout ✅

Positives:

• The empty-side check is placed before warn_mode_mismatch, so missing sessions don't also emit spurious mode warnings.
• Session IDs are {…:?}-debug-escaped in the error, preventing terminal escape-sequence injection — consistent with stats diff, and covered by missing_compare_sessions_debug_escapes_control_chars.
• Test coverage is thorough: helper unit tests (both-present / both-missing / one-side-missing / control chars), run_command tests (empty text+JSON now fail closed, one-side-missing, populated success), and binary-level integration tests against an isolated DB with numeric JSON assertions.
• Docs updated consistently in both en and zh, matching the new behavior.
• No other in-repo callers rely on the old 0% behavior.

One issue (P2, corroborates the existing bot comment)

--limit 0 produces a false-positive "No records found" — src/tokenless/crates/tokenless-cli/src/main.rs:185

The Summary --limit arg is a bare Option<usize> without the parse_positive_usize value parser that Diff.limit uses (main.rs:224). Since records_by_session applies a SQL LIMIT ?, invoking with --limit 0 returns empty vectors for populated sessions, and the new check then reports them as missing. Reproduced at head:

$ tokenless stats summary --limit 0 --compare base-run active-run   # both sessions populated
Error: No records found for baseline session "base-run" and tokenless session "active-run"
exit=1

This is exactly the failure class this PR eliminates (a misleading compare result), just gated behind an unusual flag combination. Suggested fix — reuse the existing parser, one line:

#[arg(long, value_parser = parse_positive_usize)]
limit: Option<usize>,

(Alternatively, determine session existence with an unlimited query, but the parser matches stats diff and also fixes --limit 0 for the plain summary.)

I agree with the bot's P2 on this; recommend addressing it in this PR since the change is trivial, but it is non-blocking — the common paths are correct and well covered.

Verdict: COMMENT (non-blocking). Nice work on the fail-closed alignment and the test matrix.

[Forrest-ly]

总体评价

本 PR 将 stats summary --compare 对缺失 session 的处理从「静默输出 0% 报告」改为 fail-closed（exit 1 + No records found for …），与 stats diff --session 及 Python TokenlessStats.compare 的既有语义对齐。改动聚焦、测试与文档配套完整。已在本地基于 head 45d5fe9 验证：cargo fmt --all --check、cargo clippy --workspace --locked -- -D warnings、cargo test -p tokenless-cli --locked -- --test-threads=1（264 passed, 2 ignored）全部通过；手工端到端复现确认两侧缺失 / 单侧缺失时正确失败，两侧均有记录时比较输出不受影响。

审查结论

approve

详细意见

🔴 必须修改（阻塞合并）

• 无

🟡 建议修改（不阻塞但推荐）

• [src/tokenless/crates/tokenless-cli/src/main.rs:185, 702] （补充 Codex review） 已有 Codex P2 review 指出的 --limit 0 误报问题，已在本地确认可复现：两个 session 均有记录时，tokenless stats summary --limit 0 --compare base-s act-s 报 Error: No records found for baseline session "base-s" and tokenless session "act-s"（exit 1）。根因是 Summary 的 --limit（main.rs:185）未做正数校验，records_by_session 执行 SQL LIMIT 0 返回空集，导致新增检查误判；而同文件 Diff 的 --limit 已使用 parse_positive_usize（main.rs:224、254）。建议本 PR 直接复用该 parser（或按原 review 建议独立探测 session 是否存在）。该问题已被已有 review 覆盖，此处仅为复现确认与具体修法补充，不作为新增问题、不影响本次结论；建议合并前处理。

🟢 值得肯定

• fail-closed 语义对齐到位：错误消息结构 No records found for …、exit code 1、session ID 的 debug 转义（{:?}）与 stats diff --session（main.rs:776）一致；「baseline session … and tokenless session …」的缺失侧拼接方式也与 Python 端 StatsNotFoundError（src/tokenless/python/tokenless/python/anolisa_tokenless/stats.py:377）同构，并附有控制字符注入防护的专项测试。
• 检查点放置在查询之后、warn_mode_mismatch 与格式化输出之前，保证 --json 模式下也不会泄漏 0% 报告，精确满足 issue #2673 的预期。
• 测试覆盖完整：单元测试覆盖 both-missing / baseline-only-missing / tokenless-only-missing / both-populated 四象限及控制字符转义；集成测试用隔离 TOKENLESS_STATS_DB 驱动真实二进制并断言 exit code 与 stdout/stderr；两个既有空 session 测试被正确改写为期望失败，正常路径由新增 populated 测试接管，无覆盖损失。
• 新增检查复用已查询的记录，不引入任何额外 DB 查询；en/zh 用户文档同步准确、位置恰当。

[Forrest-ly]

Re-review of head 1ceb816 (incremental diff vs previously approved 45d5fe9).

Earlier P2 (--limit 0 false missing-session report) — verified fixed:

1. stats summary --limit now uses parse_positive_usize (src/tokenless/crates/tokenless-cli/src/main.rs:184), the same parser as stats diff --limit (defined at main.rs:254). --limit 0 is rejected at clap parse time with "value must be greater than zero", so the SQL LIMIT 0 empty-result path is no longer reachable.
2. records_by_session applies LIMIT n per session (tokenless-stats/src/recorder.rs:253) and n >= 1 is now guaranteed, so an empty vector unambiguously means "no records for this session". The missing-session fail-closed contract is intact: the incremental diff touches only the arg attribute in main.rs; the missing_compare_sessions check ordering, control-char debug escaping, and exit-1 behavior are unchanged.
3. Coverage is present: parse-level stats_summary_cli_rejects_zero_limit (plain summary, summary + --compare, plus a --limit 1 positive control) and binary-level stats_summary_compare_rejects_zero_limit, which asserts populated --compare + --limit 0 fails with "greater than zero" and emits neither "No records found" nor a report.

Local gates re-run independently at head 1ceb816: cargo fmt --all --check clean; cargo clippy -p tokenless-cli --locked -- -D warnings clean; cargo test -p tokenless-cli --locked — 224 passed / 2 ignored (unit) + 42 passed (integration), 0 failed.

Note: no CI workflow runs had been triggered on this head at review time (only license/cla passing). Non-blocking for this review.

LGTM — approving.

[Forrest-ly]

Code review 通过（AI Code Review，Multica PR-Review AGE-3827，head 1ceb816）。

已在本地对 head 1ceb816 完整编译并运行 tokenless-cli 测试套件：266 tests passed / 0 failed（224 单元 + 42 集成，含本 PR 新增的 10 个 compare 用例与 2 个 zero-limit 用例）。

审查要点：

1. 正确性：missing_compare_sessions 检查位于两次 records_by_session 查询之后、任何 text/JSON 输出与 warn_mode_mismatch 之前，text 与 --json 两条路径均不会在缺 session 时输出 0% 报告，fail-closed 契约成立。--compare 的 num_args = 2 保证 sessions[0]/sessions[1] 索引安全。
2. --limit 0 误报（此前 P2 意见）：已通过复用 parse_positive_usize 在解析期拒绝，与 stats diff 一致。由于 records_by_session 使用 SQL LIMIT n（n≥1 或默认 10000），非空 session 不会再被误判为 missing。单元 + 集成测试均已覆盖并本地通过。
3. 安全性：session ID 使用 {:?} Debug 格式输出，控制字符被转义，避免构造 session ID 注入终端转义序列；与 stats diff --session 行为一致，并有专门测试验证。
4. 一致性：错误文案 No records found for ...、exit code 1 与 stats diff --session 及 CLI 现有错误约定一致；EN/ZH 文档同步更新且表述准确。

无阻塞问题，approve。

[ikunkun-sys]

审查结论

REQUEST_CHANGES

必须修改

• 本 PR 的公开 CLI 行为变更尚未完整同步到仓库要求的文档入口，详见 inline comment。

合并前提交整理

• 当前第二个提交修复的是第一个提交在本 PR 内引入的 --limit 0 边界回归。请按 AGENTS.md §13 的 same-PR fixup 规则，在合并前将两个提交 squash 为一个原子提交。

验证

• Review anchor: 1ceb816c5a2c86256f4b1696175b05deec843036
• cargo +1.89.0 fmt --all -- --check
• cargo +1.89.0 clippy --workspace --locked -- -D warnings
• cargo +1.89.0 test --workspace --locked -- --test-threads=1（590 passed，2 ignored）
• 4 个 stats summary --compare 二进制定向场景通过

远端当前仅有 license/cla 成功，尚无实际 CI workflow 结果。

[Forrest-ly]

总体评价

对 head 0a65c24 的 re-review：相对已审查通过的 1ceb816，增量改动为纯文档（6 个文件，无代码变更）。文档改动在 specs/documentation-standard.md §5 要求的两个入口（组件 README 摘要 + 用户指南完整参考）完整、准确地覆盖了本 PR 的两条公开 CLI 契约，上一轮 CHANGES_REQUESTED 中的文档缺口已解决；本地 fmt/clippy/test 全部通过。

审查结论

approve

详细意见

🔴 必须修改（阻塞合并）

• 无

🟡 建议修改（不阻塞但推荐）

• [提交历史]（补充 ikunkun-sys 既有 review）AGENTS.md §13 same-PR fixup 规则与 §6 原子性要求（"Squash fixup commits before merge"、"One commit = one logical change"）对当前 3-commit 历史仍然适用：1ceb816 修复的是 45d5fe9 在本 PR 内引入的 --limit 0 边界回归，按规则应 fixup 进源提交；新增的 0a65c24 文档化的是同一用户可见行为变更，与前两个提交共同构成一个逻辑变更，建议合并前整理为一个原子提交。作者已声明本次运行不被授权 rebase/force-push 该 PR，并提议由 GitHub squash-on-merge 达成单提交——squash-on-merge 可使 main 上得到单一原子提交，但不改写分支历史，是否满足 §13 请人类 reviewer/维护者裁定；合并前请二选一（rebase --autosquash 整理提交，或确认接受 squash-on-merge）。

🟢 值得肯定

• 两条公开契约在全部要求入口完整覆盖且中英文对齐：src/tokenless/README.md / README_zh.md 新增 stats summary 用法示例并说明两条契约；docs/user-guide/{en,zh}/token-saving/tokenless/cli-reference.md 与 measuring-savings.md 分别在 --limit 说明与 --compare 注意事项中补齐正整数契约，位置符合文档规范 §3.1/§5。
• 文档内容已逐条对照代码核实：Summary.limit 使用 parse_positive_usize（src/tokenless/crates/tokenless-cli/src/main.rs:184，--limit 0 解析期即拒绝）；missing_compare_sessions 检查位于任何 text/JSON 输出与 warn_mode_mismatch 之前（main.rs:702-709、994-1007）；records_by_session 的 LIMIT 恒为 n≥1 或默认 10,000（tokenless-stats/src/recorder.rs:189,260），空结果集必然表示该 session 无记录；"与 stats diff --limit 一致"属实（main.rs:223-224 同样使用 parse_positive_usize）。
• 全库复查未发现其他与 --compare/--limit 新行为冲突或过时的文档描述（framework-integration.md 仅涉及 Python 客户端 TokenlessStats.compare，其本就 fail-closed；user-manual.md 为泛指表述）。
• 本地门禁在 head 0a65c24 复跑：cargo fmt --all --check 通过；cargo clippy -p tokenless-cli --locked --all-targets -- -D warnings 零告警；cargo test -p tokenless-cli --locked -- --test-threads=1 → 266 passed / 2 ignored / 0 failed，含本 PR 新增的全部 compare 与 zero-limit 用例（7 个单元 + 4 个集成，与 1ceb816 轮次计数一致，确认文档提交未引入任何行为变化）。

[Forrest-ly]

Code review 通过（head 0a65c24）。本次新增 commit 为 docs-only（仅 6 个 markdown 文件），补齐了上一轮 REQUEST_CHANGES 指出的两项公开 CLI 契约文档缺口。

核对结果（docs 与 CLI 实际行为一致性）：

1. --compare 缺失 session fail-closed：在隔离环境实际运行 tokenless stats summary --compare missing-a missing-b，输出 Error: No records found for baseline session "missing-a" and tokenless session "missing-b"，exit=1，与 README（en/zh）及 user-guide measuring-savings（en/zh）新增描述一致。
2. --limit 0 解析期拒绝：实际运行 tokenless stats summary --limit 0，clap 报 invalid value '0' for '--limit <LIMIT>': value must be greater than zero，exit=2（非零），与文档"解析阶段被拒绝 / rejected at parse time"描述一致。
3. 文档中"行为与 stats diff --session / stats diff --limit 一致"的对照准确：stats diff --session 无记录时报 No records found for ...（main.rs:776），stats diff --limit 同样使用 parse_positive_usize（main.rs:224）。
4. EN/ZH 页面语义等价、命令示例完全相同，符合 specs/documentation-standard.md 与 CONTRIBUTING.md 要求；CLI flag 变更所需的文档入口（component README + user-guide 页面）均已覆盖。CHANGELOG 按仓库惯例由 release version-bump PR 汇总，本 PR 无需添加。

验证（head 0a65c24）：

• cargo test -p tokenless-cli --locked：266 passed, 2 ignored（含新增的 compare fail-closed 与 zero-limit 单元/集成测试）
• cargo fmt --all --check：clean
• cargo clippy --workspace --locked -- -D warnings：通过

代码部分（45d5fe9、1ceb816）此前已审并获批准；missing_compare_sessions 位置正确（双侧查询后、mode-mismatch 告警与报表输出前），session ID 经 {:?} debug 转义防止控制字符注入，测试覆盖双侧缺失/单侧缺失/双侧有记录（text+JSON）/zero-limit 等场景。

[zyw02]

@ikunkun-sys @Forrest-ly rebased this DIRTY/CONFLICTING PR onto current main 45306960e1dc6f1958055e97fc962b923fefb4da. New head: 72e92ed2ad094147de313ed8820b7bdcc3a4d05f (previous head 0a65c2427c09ae04d924c2ed5285f3c838902041).

Conflict file: src/tokenless/crates/tokenless-cli/tests/cli_integration.rs (imports only). Kept this PR's CompressionMode compare tests and upstream's estimate_tokens_from_bytes toon tests.

Local gates at the new head (cd src/tokenless):

• cargo fmt --all --check
• cargo clippy --workspace --locked -- -D warnings
• cargo test --workspace --locked -- --test-threads=1 (600 passed, 2 ignored)

Please re-review this revision.

[Forrest-ly]

Code review 通过（re-review @ 72e92ed）。

本轮是对 rebase 后新 head 的复核（上一轮 AGE-3841 已 APPROVE 0a65c24），重点核对 rebase 增量与遗留线程：

1. Rebase 增量无行为变化（已逐 hunk 核对）
对比 9f34d03..0a65c24（旧 patch）与 45306960..72e92ed（新 patch）：除 tests/cli_integration.rs 的 import hunk 外，所有 hunk 完全一致。该 hunk 即 PR 描述中声明的冲突解决——同时保留本 PR 的 CompressionMode 与上游的 estimate_tokens_from_bytes，无其他改动。

2. 核心修复复核（src/tokenless/crates/tokenless-cli/src/main.rs）

• missing_compare_sessions 检查位于两侧查询之后、warn_mode_mismatch 与任何报告输出之前，fail-closed（exit 1）时序正确。
• 错误信息 No records found for ... 前缀与 {sid:?} debug 转义均与 stats diff --session（main.rs:776）保持一致，控制字符注入有单测覆盖。
• Summary.limit 改用共享的 parse_positive_usize（main.rs:254），与 stats diff --limit 一致，--limit 0 在解析期即被拒绝（"value must be greater than zero"），消除了 limit=0 伪装成 session 缺失的歧义。

3. 两条遗留 P2 线程在当前 head 已确认解决

• chatgpt-codex-connector 的 --limit 0 P2：由 c015989 修复，stats_summary_cli_rejects_zero_limit / stats_summary_compare_rejects_zero_limit 覆盖（含 populated compare 场景）。
• ikunkun-sys 的文档 P2：README（en/zh）与 user-guide（en/zh）四处均已写明两项公开契约（缺失 session fail-closed、--limit 必须为正整数），与 CLI 实际行为一致。

4. 本地验证（head 72e92ed）

• cargo test -p tokenless-cli --locked：275 passed, 2 ignored（224 unit + 46 integration + 5 matcher contract），与 PR 描述一致；新增/更新的 7 个 compare 相关单测全部通过。
• cargo fmt --all --check 通过。

既有测试中 run_command_stats_compare / run_command_stats_compare_json 由 is_ok() 改为 fail-closed 断言，属于本次有意的契约变更，改法正确。无新增问题，approve。

[ikunkun-sys]

审查结论

REQUEST_CHANGES

必须修改：与最新 main 存在代码冲突

当前 head 72e92ed2ad094147de313ed8820b7bdcc3a4d05f 无法合并到最新 main ec7c99ef2887d67a7e49d176c9ef319b47e97d48；GitHub 当前状态为 CONFLICTING / DIRTY。

三方合并证据显示，实际未解决的文本冲突在 src/tokenless/crates/tokenless-cli/src/tests/main_tests.rs：

• 上游 f79e8300 在同一测试区域新增了 stats_persist_snapshot_enable_keeps_file_compression_and_sls / stats_persist_snapshot_disable_keeps_file_compression_and_sls。
• 本 PR 在该区域新增了 missing_compare_sessions_* 测试和 seed_compare_record helper。

src/tokenless/crates/tokenless-cli/src/main.rs 也由双方修改，但当前三方合并可自动处理；真正产生 conflict markers 的是上述 main_tests.rs 测试块。请 rebase 到最新 main，保留两组测试及对应实现，然后重新运行 Tokenless 的 fmt、Clippy 和 workspace tests。

提交历史

当前分支仍有 3 个提交。处理 rebase 冲突时，请一并按此前 review 和 AGENTS.md §13 的 same-PR fixup 规则，将它们整理为一个原子提交。

CI 状态

当前 Tokenless、Docs、Pages 和 PR lint checks 均为成功，但这些结果不能消除当前分支与最新 main 的合并冲突。

[zyw02]

@ikunkun-sys rebased this DIRTY/CONFLICTING PR onto current main a4cb23ea27c9645d7294b7afe65309262e6a0e4b. New head: 6278f1db9cfcff64bb8cae83ff81009e09664410 (previous head 72e92ed2ad094147de313ed8820b7bdcc3a4d05f).

Conflict file: src/tokenless/crates/tokenless-cli/src/tests/main_tests.rs. Kept upstream persist snapshot / enable-disable tests and this PR's missing_compare_sessions_* coverage plus seed_compare_record. main.rs auto-merged; upstream persist_stats_enabled safeguards remain.

The previous 3 commits were squashed into one atomic commit on this rebase.

Local gates at the new head (cd src/tokenless):

• cargo fmt --all --check
• cargo clippy --workspace --locked -- -D warnings
• cargo test --workspace --locked -- --test-threads=1 (606 passed, 2 ignored)

Please re-review this revision.

[Forrest-ly]

Re-review APPROVE @ 6278f1d（rebase onto latest main a4cb23ea，单 squashed commit）

Rebase 增量复核（重点）：逐 hunk 对比了上一轮已批准的累计 diff（45306960..72e92ed，3 commits）与本轮累计 diff（a4cb23ea..6278f1d，1 commit）。两者 PR 内容逐行一致（643 行 diff 完全等长），差异仅为：

1. blob hash 与 hunk 行号偏移（base 变化所致，预期内）；
2. main_tests.rs 中本 PR 测试块的插入点上移：从 run_command_stats_status 之后移到 upstream 新增的 run_command_stats_disable_does_not_persist_env_overrides（main_tests.rs:1218）之后，@@ -1072,26 +1104,174 @@ → @@ -1197,26 +1229,174 @@，hunk 尺寸一致（+174），无任何内容行变化。

冲突解决确认：main_tests.rs（唯一冲突文件）同时保留了 upstream 的 stats_persist_snapshot_* / run_command_stats_{enable,disable}_does_not_persist_env_overrides 测试与本 PR 的 missing_compare_sessions_*、seed_compare_record、stats_summary_cli_rejects_zero_limit 测试；main.rs 中 upstream 的 persist_stats_enabled / stats_persist_snapshot 路径（enable/disable 命令分支）与本 PR 的 fail-closed --compare（summary 命令分支 main.rs:699）及 parse_positive_usize（main.rs:184）互不干扰。

本地验证（exact head 6278f1db）：

• cargo fmt --all --check ✅
• cargo clippy --workspace --locked -- -D warnings ✅
• cargo test --workspace --locked -- --test-threads=1 ✅ 606 passed, 2 ignored（与 PR 描述一致）

结论：rebase 增量无行为变化，维持此前 APPROVE 的全部结论（fail-closed --compare、--limit 0 解析期拒绝、文档与回归测试）。