[code-review] #142 fix in unchecked_assert_check.go/return_count_check.go uses position strings as 'content fingerprints' — line shifts re-flag existing issues

[topcheer]

File/Line

• internal/agent/unchecked_assert_check.go L171-184 (findUncheckedAsserts / assertFingerprint)
• internal/agent/return_count_check.go L51-95 (checkExcessiveReturns delta logic)

Problem

The #142 fix replaced count-based deltas with set-based fingerprints, but the fingerprint is token.Position.String() — a file:line:col location string, not content:

// unchecked_assert_check.go L174
results = append(results, uncheckedAssertInfo{line: p.Line, expr: p.String()}) // expr = "/path/file.go:42:1"

The fix's own comment claims the opposite of the behavior: "Uses the expression text so the same assertion at different line numbers is recognized as the same issue" — with a position key, the same assertion at a different line is always a NEW issue.

Trigger scenario

1. helper() at L100 has 8 returns (pre-existing, over threshold); two unchecked assertions exist at L10/L15
2. Agent inserts 5 lines at the top of the file for an unrelated fix
3. Old fingerprint file.go:100:1 vs new file.go:105:1 — mismatch
4. The untouched function/assertions are flagged as newly-introduced → false warnings

Net effect: any edit above existing problem sites re-flags every pre-existing issue below it — the same false-positive class #142 was meant to eliminate, relocated from count drift to line drift.

Expected vs actual

• Expected: unchanged issues survive line shifts (comment's stated intent)
• Actual: line shift = new fingerprint = false "introduced" warning

Fix

• unchecked_assert: fingerprint on the assertion expression text (e.g. x.(int))
• return_count: fingerprint on the function name
• Correct or delete the misleading comment

Severity

High — deterministic false positives on a routine edit pattern (inserting lines above existing code); fires on every write-integrity pass.

Verified by independent review subagent (sa-3) with concrete old/new interleaving; ctxkey_check.go compared as reference.

[topcheer]

已修复并推送（commit 14a4c14）。

根因

#142 的 set-based delta 用 token.Position.String()（file:line:col 位置字符串）作为内容指纹。注释声称"同一断言在不同行号应识别为同一 issue"，但位置 key 使行号一变即视为新问题——在问题点上方的任何编辑都会重新标记下方所有存量问题，假阳性只是从计数漂移转移到了行漂移。

修复

• unchecked_assert_check.go：指纹改为断言表达式文本。新增 assertExprAt（AST 定位 + 复用包级 exprText 渲染），如 v.(int)，与行号无关。
• return_count_check.go：指纹改为函数名（oldPos[iss.funcName]），行偏移不影响匹配。

测试

• TestCheckUncheckedTypeAssert_LineShiftNotReflagged：两个存量断言上方插入 5 行无关注释，验证不再误报
• TestCheckExcessiveReturns_LineShiftNotReflagged：同场景验证 return-count 不再误报

internal/agent 全包测试通过。