perf(contract): avoid resolving unsupported regular scan files

docs/architecture/rfcs/typescript-control-plane-migration-v0.md

@@ -2236,6 +2236,48 @@ The same real File/SQLite entrypoints must qualify functional parity and the
 original full-CLI cost gate together. Fewer crossings alone do not explain a
 whole-CLI tail regression or complete writer/backup/default-provider retirement.
 
+Fresh public-boundary scanning retains its specialized Python regex / Host IO
+owner. Characterization covers all eleven `splitlines`
+separators, Unicode regex folding, casefold expansion false positives, ordered
+multi-rule hits, credential references, the public-host exception and literal
+metacharacters. These are scanner semantics, not a new decision owner or cache.
+
+A folded-literal union intervention reduced the exact quota scan population's
+median wall / CPU cost by 8.5% / 6.9% with complete scanner payload parity.
+Its fixed 32 alternating full-CLI pairs per provider still produced File p95
+976.33 → 986.84 ms (+10.52 ms / 1.08%) and SQLite p95 936.00 → 962.34 ms
+(+26.33 ms / 2.81%). The latter misses the unchanged unexplained 25 ms target,
+so the runtime candidate was withdrawn. All samples remain evidence; profiler
+observations do not establish the tail cause or waive the qualification.
+Further full-payload controls rejected folded-line reuse and literal regex
+disjunctions; composing the existing direct-loop and line-union interventions
+improved scanner wall time by only 3.6%. An independent fixed 32-pair A/A control
+per provider, with identical runtime bytes and the same complete CLI workload,
+still differed by -8.20 ms / -0.84% File p95 and -85.95 ms / -9.90% SQLite p95
+between checkout contexts. All 128 commands retained the expected selection and
+frontier. This identifies a comparison confound, not its cause or an optimizer
+qualification; the earlier failed candidate and all observations remain.
+A same-checkout cold/warm control then isolated runtime startup: all 32 cold
+samples changed the setup PID and were slower than their warm partner; warm
+samples retained substantial variability. This explains a startup contribution,
+not the historical A/A tail difference.
+
+The next intervention rejects unsupported regular filenames before repeated
+canonical resolution in both Git enumeration and directory traversal. Symlinks
+still resolve before suffix/local-name eligibility; tracked pruned-directory
+files, explicit file roots and fresh retargets retain their original behavior.
+On the unchanged package-root workload it removed 602 of 3,281 resolution calls
+with full payload parity; eight alternating scan pairs improved median wall
+time by 5.1%. Fixed 32-pair full quota CLI comparisons per provider used one
+physical checkout and verified warm setup PID reuse in every sample. File
+p50/p95 was 708.70/891.78 → 689.04/856.32 ms; SQLite was
+758.55/1036.27 → 769.41/1026.34 ms. Both pass the original additive 25 ms and
+5% tail budgets, while SQLite's slightly higher median limits the improvement
+claim. This qualifies this enumeration intervention, not the withdrawn filter,
+the historical frontier/direct-loop holds or a frozen multi-command deadline.
+Preserve those separately from new-Goal defaults, recoverable imports, the
+developer trial and last-caller retirement.
+
 2026-09-24: [Typed complete-source assembly and remaining delivery packages](ledger/shared-goal-authority-state-provider-v0/2026-09-24-source-capture.md) unify source construction, identity rejection and current-graph membership; L7/D2/D3 and provider defaults remain open.
 
 2026-09-24: [Leased continuation and remaining local-default packages](ledger/shared-goal-authority-state-provider-v0/2026-09-24-leased-continuation.md).

docs/architecture/rfcs/typescript-control-plane-migration-v0.zh-CN.md

@@ -1630,6 +1630,36 @@ TS 摘要批次；Python 保留旧格式解码、公开字段筛选及渲染。
 查询命中情况覆盖。见[语义及回滚](../../reference/todo-work-counts.md)。这是 T3/L5 的
 共享读取边界推进，不替代 D2/D3 或 provider 默认切换。
 
+fresh 公共边界扫描继续归现有专门 Python regex／Host IO owner。
+Characterization 已覆盖全部十一种 `splitlines` 分隔符、Unicode regex 折叠、
+casefold 扩展的假阳性、逐规则命中顺序、凭据引用、公共主机例外及字面量元字符；
+这些是 scanner 语义，不新增决策 owner 或缓存。
+
+折叠字面量合并的受控干预，在真实 quota 扫描集合上降低中位 wall／CPU 成本
+8.5%／6.9%，完整 scanner payload 相等。但每 provider 固定 32 对交替完整 CLI
+比较仍得到 File p95 976.33 → 986.84 ms（+10.52 ms／1.08%）、SQLite p95
+936.00 → 962.34 ms（+26.33 ms／2.81%）。后者未通过既有未解释 25 ms 目标，
+因此撤回运行时候选，保留全部样本；profile 不能证明尾延迟根因或豁免资格。
+后续完整 payload 对照淘汰了 folded-line 复用及 literal regex 合并；既有
+direct-loop 与 line-union 组合仅降低 scanner wall 3.6%。另一次独立固定每 provider
+32 对 A/A 对照，在运行时字节及完整 CLI 负载相同的两个 checkout 间，仍得到
+File p95 -8.20 ms／-0.84%、SQLite p95 -85.95 ms／-9.90%；全部 128 次命令的
+selection／frontier 均符合预期。这证明比较存在混杂，尚未证明其根因或优化资格；
+此前候选失败及全部观察继续保留。同一 checkout 的冷暖对照随后隔离了 runtime
+启动：32 次冷态均更换 setup PID，且比暖态配对更慢；暖态仍明显波动。
+这证明启动成本贡献，未解释历史 A/A 尾差。
+
+下一项干预在 Git 枚举和目录遍历中，先跳过后缀不支持的普通文件，再做 canonical
+路径解析；符号链接仍先解析目标，再判断后缀及 local 名称。Git 跟踪的剪枝目录、
+显式文件 root 及链接改指向后的 fresh 读取保持原行为。同一 package-root 负载少做
+602／3,281 次解析，完整 payload 相等；8 对交替 scanner 中位 wall 改善 5.1%。
+每 provider 固定 32 对完整 quota CLI，在同一 physical checkout 逐次确认暖态
+setup PID 复用。File p50/p95 为 708.70/891.78 → 689.04/856.32 ms；SQLite 为
+758.55/1036.27 → 769.41/1026.34 ms，均通过原 25 ms 及 5% 尾延迟预算。
+SQLite 中位数略升，收益不能外推为全面加速。这只资格化本次枚举干预，不改写
+已撤回 filter 的失败，也不解除历史 frontier／direct-loop hold 或冻结多命令时限。
+这些与新 Goal 默认、可恢复导入、开发者试用及最后调用方退役分别保留。
+
 2026-09-24：[完整源捕获的 TS 组装与剩余交付包](ledger/shared-goal-authority-state-provider-v0/2026-09-24-source-capture.zh-CN.md)统一源构造、身份拒绝和当前图成员规则；不关闭 L7/D2/D3 或启用默认 provider。
 
 2026-09-24: [带租约接力与剩余本地默认交付包](ledger/shared-goal-authority-state-provider-v0/2026-09-24-leased-continuation.zh-CN.md).

loopx/contract.py

@@ -847,7 +847,10 @@ def _tracked_scan_files(scan_root: Path) -> list[Path]:
     for rel_path in tracked.stdout.split("\0"):
         if not rel_path:
             continue
-        path = (repo_root / rel_path).resolve()
+        path = repo_root / rel_path
+        if path.suffix not in DEFAULT_SCAN_SUFFIXES and not path.is_symlink():
+            continue
+        path = path.resolve()
         if path.name.endswith(".local.json"):
             continue
         if path.is_file() and path.suffix in DEFAULT_SCAN_SUFFIXES:
@@ -879,7 +882,10 @@ def iter_scan_files(scan_root: Path) -> list[Path]:
             and not os.path.isfile(current_dir / name / "pyvenv.cfg")
         ]
         for file_name in file_names:
-            path = (current_dir / file_name).resolve()
+            path = current_dir / file_name
+            if path.suffix not in DEFAULT_SCAN_SUFFIXES and not path.is_symlink():
+                continue
+            path = path.resolve()
             if path.name.endswith(".local.json"):
                 continue
             if not path.is_file():

loopx/semantics/project_registry_io_manifest_v1.json

@@ -1039,7 +1039,7 @@
     },
     {
       "site": "loopx/contract.py::<module>.check_contract::codec_read:load_registry#1",
-      "line": 1093,
+      "line": 1099,
       "column": 20,
       "kind": "codec_read",
       "api": "load_registry",

tests/test_contract_public_boundary_prefilter.py

@@ -1,13 +1,62 @@
 from __future__ import annotations
 
 from pathlib import Path
+import subprocess
 from typing import Any
 
 import pytest
 
 from loopx import contract
 
 
+@pytest.mark.parametrize("tracked,pruned", [(False, False), (True, False), (True, True)])
+def test_scan_enumeration_uses_resolved_symlink_suffix_and_local_name(
+    tmp_path: Path, tracked: bool, pruned: bool
+) -> None:
+    scan_root = tmp_path / ("node_modules" if pruned else "public")
+    scan_root.mkdir()
+    target = tmp_path / "target.md"
+    target.write_text("public", encoding="utf-8")
+    distinct = tmp_path / "distinct.md"
+    distinct.write_text("public", encoding="utf-8")
+    unsupported = tmp_path / "target.ts"
+    unsupported.write_text("not a directory scan input", encoding="utf-8")
+    local = tmp_path / "target.local.json"
+    local.write_text("local", encoding="utf-8")
+    regular = scan_root / "regular.md"
+    regular.write_text("public", encoding="utf-8")
+    (scan_root / "regular.ts").write_text("unsupported", encoding="utf-8")
+    (scan_root / "alias.ts").symlink_to(target)
+    (scan_root / "distinct.txt").symlink_to(distinct)
+    (scan_root / "alias.local.json").symlink_to(target)
+    (scan_root / "unsupported.md").symlink_to(unsupported)
+    (scan_root / "local.md").symlink_to(local)
+    (scan_root / "broken.ts").symlink_to(tmp_path / "absent.md")
+    if tracked:
+        subprocess.run(["git", "init", "-q", str(tmp_path)], check=True)
+        subprocess.run(["git", "-C", str(tmp_path), "add", "-f", "--", scan_root.name], check=True)
+
+    # Git ownership includes tracked files under otherwise pruned directories.
+    # Directory eligibility follows the canonical target, not the alias name.
+    # The pruned-root fast path preserves Git's per-alias observations.
+    expected = [distinct, regular, target, target] if pruned else [distinct, regular, target]
+    assert contract.iter_scan_files(scan_root) == sorted(expected)
+    if tracked:
+        assert set(contract._tracked_scan_files(scan_root)) == {distinct, regular, target}
+
+    # A later retarget must be observed; no eligibility result may be cached.
+    (scan_root / "alias.ts").unlink()
+    (scan_root / "alias.ts").symlink_to(local)
+    (scan_root / "alias.local.json").unlink()
+    assert contract.iter_scan_files(scan_root) == sorted([distinct, regular])
+
+
+def test_explicit_scan_file_keeps_unsupported_suffix(tmp_path: Path) -> None:
+    explicit = tmp_path / "explicit.ts"
+    explicit.write_text("explicit input", encoding="utf-8")
+    assert contract.iter_scan_files(explicit) == [explicit]
+
+
 @pytest.mark.parametrize(
     ("rule_name", "line"),
     [
@@ -110,3 +159,69 @@ def test_prefilter_preserves_all_boundary_hit_categories(tmp_path: Path) -> None
         "sample.md:4: internal_task_id",
         "sample.md:5: private_ip",
     ]
+
+
+@pytest.mark.parametrize(
+    "separator",
+    [
+        "\n",
+        "\r",
+        "\r\n",
+        "\v",
+        "\f",
+        "\x1c",
+        "\x1d",
+        "\x1e",
+        "\x85",
+        "\u2028",
+        "\u2029",
+    ],
+)
+def test_prefilter_preserves_unicode_line_numbers_and_multiple_rules(
+    tmp_path: Path, separator: str
+) -> None:
+    # The folded necessary condition cannot change original line coordinates,
+    # rule order, the public-host exception, or credential-reference handling.
+    lines = [
+        "ordinary " * 50,
+        "Author\u0131zation: literal host 10" + ".1.2.3",
+        "pa\u017f\u017fword=${EXAMPLE_KEY}",
+        "https://open.lark" + "office.com host 172" + ".31.2.3",
+        "pa\u00dfword=literal",  # casefold expands; the authoritative regex refuses it
+        "ordinary again",
+    ]
+    (tmp_path / "sample.md").write_text(separator.join(lines), encoding="utf-8")
+    payload = contract.scan_public_boundary([tmp_path])
+    assert payload["hits"] == [
+        "sample.md:2: credential",
+        "sample.md:2: private_ip",
+        "sample.md:4: private_ip",
+    ]
+    assert payload["credential_reference_hits"] == ["sample.md:3: credential"]
+
+
+def test_prefilter_literals_are_substrings_not_regex_syntax(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    class RecordingPattern:
+        def __init__(self) -> None:
+            self.lines: list[str] = []
+
+        def search(self, line: str) -> None:
+            self.lines.append(line)
+
+    pattern = RecordingPattern()
+    monkeypatch.setattr(
+        contract,
+        "LEAK_RULES",
+        {
+            "private_ip": contract.LeakRule(
+                pattern=pattern,  # type: ignore[arg-type]
+                required_literals=("a|b", "10."),
+            ),
+        },
+    )
+    lines = ["a or b", "10x", "actual a|b", "actual 10.", "ordinary"]
+    (tmp_path / "sample.md").write_text("\n".join(lines), encoding="utf-8")
+    assert contract.scan_public_boundary([tmp_path])["ok"] is True
+    assert pattern.lines == ["actual a|b", "actual 10."]