diff --git a/.agents/issues/BACKEND-VULKAN/ISSUE-LOCAL-01M3JJ9CE2WH0KE640HQFGACNH.md b/.agents/issues/BACKEND-VULKAN/ISSUE-LOCAL-01M3JJ9CE2WH0KE640HQFGACNH.md
new file mode 100644
index 0000000000..db7e10decf
--- /dev/null
+++ b/.agents/issues/BACKEND-VULKAN/ISSUE-LOCAL-01M3JJ9CE2WH0KE640HQFGACNH.md
@@ -0,0 +1,32 @@
+ID: ISSUE-LOCAL-01M3JJ9CE2WH0KE640HQFGACNH
+Title: No .gitattributes line-ending policy: agent-issue-index.py --refresh can never pass on a CRLF checkout
+Row: BACKEND-VULKAN
+State: OPEN
+Kind: bug
+GitHub: -
+Mirror: PENDING
+Availability: FULL
+Created: 2026-09-27
+Updated: 2026-10-02
+Closed: -
+
+## Problem
+
+The repository has NO line-ending policy. .gitattributes exists but contains only a linguist-generated rule for the vendored Triton AOT artifacts, so on any contributor machine with core.autocrlf=true (the default on Windows) every text file is checked out CRLF.
+
+Several tools compare file content as EXACT BYTES, and one of them is now structurally unable to pass on such a checkout. scripts/issue_records.py _archive_evidence_matches_source splits the frozen archive on b newline and requires the raw bytes at a declared line to equal the evidence quoted in an _intake issue. parse_issue_file normalises the quoted evidence through normalize_body, which runs _normal_newlines and strips the carriage return. So on a CRLF checkout the archive line ends CR and the quote does not, and they can never be equal.
+
+MEASURED 2026-09-27 on a Windows checkout at 1de097c46, for ISSUE-GH-148: the archive line and the quoted evidence are byte-identical except for one trailing CR, EXACT MATCH False, match-ignoring-CR True. The committed blob is pure LF (0 CR, 886 LF in .agents/completed/issue-index.md), so the repository content is correct and this is purely a checkout artifact. Consequence: scripts/agent-issue-index.py --refresh fails permanently for Windows contributors, which silently disables the canonical issue index, and the failure reads like a record defect rather than an environment one.
+
+The fix is the line endings, NOT the checker. Making the comparison CR-tolerant would widen a checker to make a gate green, which AGENTS.md forbids without a spec and red-before evidence, and byte-exactness is the invariant the record actually depends on.
+
+SCOPE REVISED 2026-10-02 after review asked for the checkout test. The first cut covered only .agents/**/*.md, scripts/*.py and the DeepSeek-V4-Vision fixture dir, reasoning that those were the paths a byte-exact checker reads. A core.autocrlf=true clone + checkout of the branch showed that framing was wrong in BOTH directions:
+
+- The record surface is wider than *.md. On the autocrlf checkout the .md rows stayed LF, but .agents/completed/*.csv manifests, .agents/evidence/** JSON/stdout captures, .agents/specs/*.log and *.patch, and .agents/scripts/*.sh / *.py all arrived CRLF -- the same byte-exact hazard one directory over.
+- The policy normalises nothing that exists. Scanning the index finds ZERO text blobs outside docs/bench-evidence containing \r\n, and ZERO binary blobs under .agents/ (2488 files) or scripts/ (313), so a whole-tree * text=auto eol=lf is safe and rewrites no stored content -- it only stops the class from arriving.
+
+The policy is therefore `* text=auto eol=lf` with two byte-preserving pins. docs/bench-evidence/** is pinned -text directory-wide, not file-by-file: the directory carries three true-CRLF text logs (job-phase2.txt, profile-dependencies.log.txt, gen-20260903T012806Z.log) PLUS eight more text logs with lone progress-bar CRs (limb3-strict-gate gen-*.log, job-phase3.txt, q4km-neartie harness-stdout-*.txt), and any of them committed from an autocrlf checkout would be rewritten. tests/parity/goldens/** is pinned -text because the .npy/.i32/.raw fixtures and generated manifests are bytes, not lines. After the change, the same autocrlf clone produces zero CR-bearing files outside the pinned evidence directory.
+
+## Resolution
+
+-
diff --git a/.gitattributes b/.gitattributes
index d67e7e7212..31842d22b3 100644
--- a/.gitattributes
+++ b/.gitattributes
@@ -1,3 +1,38 @@
 # Vendored Triton AOT artifacts are generated code (embedded cubins): collapse
 # in diffs/review and exclude from language stats. Regen: scripts/regen-triton-aot.sh
 src/vt/cuda/triton_aot_vendored/** linguist-generated=true
+
+# ---------------------------------------------------------------------------
+# Line endings.
+#
+# Several checkers compare file content as EXACT BYTES, so a CRLF checkout
+# makes them fail on content that is correct. The worst case is
+# `_archive_evidence_matches_source` in scripts/issue_records.py: it splits the
+# frozen archive on a newline byte and requires the raw bytes at a declared
+# line to equal evidence quoted in an _intake issue, but that quote is
+# normalised through `normalize_body`, which strips the carriage return. On a
+# CRLF checkout the two differ by exactly one CR, so
+# `agent-issue-index.py --refresh` can never pass -- permanently, and for a
+# reason that reads like a bad record rather than a bad environment.
+#
+# Measured scope, not a guess: a `core.autocrlf=true` clone + checkout of this
+# revision was compared against the LF worktree. The record surface is wider
+# than *.md -- .agents/completed/*.csv manifests, .agents/evidence/** JSON and
+# stdout captures, .agents/specs/*.log and *.patch, and .agents/scripts/*.sh /
+# *.py all arrived CRLF while the .md rows stayed LF. There are ZERO binary
+# blobs under .agents/ or scripts/ (2488 and 313 files scanned), so a
+# directory-wide text rule is safe; and there are ZERO tracked text blobs
+# outside docs/bench-evidence that contain \r\n, so this policy normalises
+# nothing that exists today -- it only stops the class from arriving.
+* text=auto eol=lf
+
+# Captured bench evidence, where a CR is part of the recorded artifact and must
+# survive verbatim. Pinned directory-wide rather than file-by-file: the tree
+# carries more CR-bearing text logs than the three files that contain \r\n
+# (progress-bar carriage returns in limb3-strict-gate and q4km-neartie logs),
+# and any of them committed from an autocrlf checkout would be rewritten.
+docs/bench-evidence/** -text
+
+# Parity goldens are fixture bytes -- .npy/.i32/.raw tensors and generated
+# manifests -- never line-ending-normalised.
+tests/parity/goldens/** -text