diff --git a/crates/unsafe-review-core/src/analysis/scanner.rs b/crates/unsafe-review-core/src/analysis/scanner.rs
index 21e938801..493162bc1 100644
--- a/crates/unsafe-review-core/src/analysis/scanner.rs
+++ b/crates/unsafe-review-core/src/analysis/scanner.rs
@@ -737,6 +737,9 @@ fn is_public_api_surface(kind: &UnsafeSiteKind, snippet: &str) -> bool {
 fn find_owner(lines: &[&str], idx: usize) -> Option<String> {
     for raw in lines[..=idx].iter().rev().take(80) {
         let line = raw.trim();
+        if is_comment_line(line) {
+            continue;
+        }
         if let Some(name) = parse_impl_owner(line) {
             return Some(name);
         }
@@ -765,6 +768,9 @@ fn find_owner_declaration_index(lines: &[&str], idx: usize) -> Option<usize> {
     let limit = idx.min(lines.len().saturating_sub(1));
     for (line_idx, raw) in lines[..=limit].iter().enumerate().rev().take(120) {
         let line = raw.trim();
+        if is_comment_line(line) {
+            continue;
+        }
         if parse_impl_owner(line).is_some()
             || parse_trait_name(line).is_some()
             || parse_fn_name(line).is_some()
@@ -775,6 +781,10 @@ fn find_owner_declaration_index(lines: &[&str], idx: usize) -> Option<usize> {
     None
 }
 
+fn is_comment_line(line: &str) -> bool {
+    line.starts_with("//") || line.starts_with("/*") || line.starts_with('*')
+}
+
 fn owner_doc_start(lines: &[&str], decl_idx: usize) -> usize {
     let mut start = decl_idx;
     let mut idx = decl_idx;
@@ -918,6 +928,20 @@ mod tests {
         );
     }
 
+    #[test]
+    fn owner_inference_ignores_comment_text_when_scanning_backwards() {
+        let lines = [
+            "fn keep_rest(&mut self) {",
+            "    unsafe {",
+            "        // Normally `Drop` impl would drop [tail].",
+            "        let src = ptr.add(this.idx);",
+            "    }",
+            "}",
+        ];
+
+        assert_eq!(find_owner(&lines, 3), Some("keep_rest".to_string()));
+    }
+
     #[test]
     fn text_detection_does_not_classify_deref_assignments_as_writes() {
         assert_eq!(detect_site("*ptr = value;"), None);
diff --git a/docs/handoffs/2026-05-18-real-crate-dogfood-v0.6.md b/docs/handoffs/2026-05-18-real-crate-dogfood-v0.6.md
index a2fa00261..e9060c1f1 100644
--- a/docs/handoffs/2026-05-18-real-crate-dogfood-v0.6.md
+++ b/docs/handoffs/2026-05-18-real-crate-dogfood-v0.6.md
@@ -67,6 +67,17 @@ evidence-quality gap:
 Regression proof was added with a repo-mode analyzer test for pointer
 arithmetic inside a documented unsafe function.
 
+A real PR-diff dogfood pass on `servo/rust-smallvec#407` exposed and fixed an
+owner-inference gap:
+
+- owner inference now ignores comment lines while scanning backward
+- comment prose such as `` `Drop` impl would drop `` is no longer parsed as an
+  `impl would` owner
+- witness commands now use the real owner when one is available
+
+Regression proof was added with a scanner unit test for comment text during
+owner inference.
+
 ## Dogfood observations
 
 The before/after numbers below are top-50 capped repo inventory snapshots, not
@@ -208,6 +219,45 @@ The pointer arithmetic card now recognizes the enclosing `# Safety` contract and
 local guard evidence, then routes the remaining gap to Miri/cargo-careful as a
 witness need.
 
+### `rust-smallvec#407`
+
+PR: `https://github.com/servo/rust-smallvec/pull/407`
+
+The PR fixes a use-after-free in `DrainFilter::keep_rest` for zero-capacity
+`SmallVec`s by changing the ZST guard and adding a Miri-targeted regression
+test.
+
+Before comment-aware owner inference:
+
+```text
+elapsed_seconds: 33.48
+changed_rust_files: 2
+cards: 1
+contract_missing: 1
+owner: would
+verify: cargo +nightly miri test would
+```
+
+The owner was inferred from comment prose:
+
+```text
+// Normally `Drop` impl would drop [tail] ...
+```
+
+After comment-aware owner inference:
+
+```text
+elapsed_seconds: 34.8
+changed_rust_files: 2
+cards: 1
+contract_missing: 1
+owner: keep_rest
+verify: cargo +nightly miri test keep_rest
+```
+
+The card class did not change. The improvement is that the card now points the
+reviewer and witness route at the real owner instead of prose.
+
 ## Proof
 
 Targeted local validation:
@@ -229,6 +279,7 @@ rtk cargo run --locked -p unsafe-review -- repo --root target/dogfood-work/small
 rtk cargo run --locked -p unsafe-review -- repo --root target/dogfood-work/arrayvec --format json --max-cards 50 --out target/dogfood-work/arrayvec.unsafe-review.after.json
 rtk cargo run --locked -p unsafe-review -- repo --root target/dogfood-work/memchr --format json --max-cards 50 --out target/dogfood-work/memchr.unsafe-review.after-cap-targetfeature.json
 rtk cargo run --locked -p unsafe-review -- check --root target/dogfood-work/memchr --diff target/dogfood-work/memchr-pr215.raw.diff --format json --max-cards 20 --out target/dogfood-work/memchr-pr215.owner-contract.json
+rtk cargo run --locked -p unsafe-review -- check --root target/dogfood-work/smallvec --diff target/dogfood-work/smallvec-pr407.raw.diff --format json --max-cards 20 --out target/dogfood-work/smallvec-pr407.owner-fix.json
 ```
 
 The dogfood reruns used a temporary `CARGO_TARGET_DIR` to avoid a Windows file
@@ -242,11 +293,13 @@ The repo may claim:
 
 - the first real-crate dogfood slice was run on `rust-smallvec` and `arrayvec`
 - a capped `memchr` dogfood snapshot now completes
-- a real PR-diff dogfood run on `memchr#215` produces card output
+- real PR-diff dogfood runs on `memchr#215` and `rust-smallvec#407` produce card
+  output
 - dogfood found and fixed import/declaration and `cfg(target_feature)`
   false positives
 - capped repo scans stop after the requested card cap
 - operation cards can inherit enclosing unsafe function `# Safety` docs
+- owner inference ignores comments while scanning backward
 - false-positive regression coverage exists in fixtures and calibration
 - dogfood output remains advisory static review evidence
 
@@ -256,7 +309,7 @@ The repo must not claim:
 - usable-alpha support-tier promotion
 - full-repository coverage from top-50 capped snapshots
 - uncapped repo-scan performance
-- general PR-diff usefulness from one PR
+- general PR-diff usefulness from two PRs
 - memory-safety proof
 - UB-free status
 - witness execution
@@ -266,7 +319,7 @@ The repo must not claim:
 
 - Only three real crates completed in this slice.
 - The successful dogfood snapshots were capped at 50 cards.
-- Only one real PR diff was measured.
+- Only two real PR diffs were measured.
 - `memchr` completion depends on capped-scan behavior; uncapped performance is
   still unmeasured.
 - No human audit was performed for every emitted card.
diff --git a/docs/status/OBJECTIVE_AUDIT.md b/docs/status/OBJECTIVE_AUDIT.md
index bf9a2ba98..c83550e67 100644
--- a/docs/status/OBJECTIVE_AUDIT.md
+++ b/docs/status/OBJECTIVE_AUDIT.md
@@ -35,7 +35,7 @@ witnesses by default.
 | Explicit receipts can be authored and validated safely | `receipt template` and `receipt validate` are covered by CLI e2e tests and support tiers | Experimental | Template output does not verify that the recorded command ran |
 | Public claims map to proof | `SUPPORT_TIERS.md` maps every current surface to proof and limits | In place | Keep updating for every new lane |
 | No soundness, UB-free, Miri-clean, site-execution, or default-blocking claim | Trust-boundary text is enforced across artifacts; support tiers and handoffs repeat limits | In place | Must remain part of all new projections |
-| First real-crate dogfood measurement | Handoff `2026-05-18-real-crate-dogfood-v0.6.md` records top-50 capped `rust-smallvec`, `arrayvec`, and `memchr` runs plus a `memchr#215` PR-diff run; dogfood found and fixed import/declaration false positives, `cfg(target_feature)` false positives, capped repo scan timeout behavior, and missing owner-contract inheritance for operation cards | Experimental | More crates, more real PR diffs, uncapped/sampled runs, and human review are still needed before calibration claims |
+| First real-crate dogfood measurement | Handoff `2026-05-18-real-crate-dogfood-v0.6.md` records top-50 capped `rust-smallvec`, `arrayvec`, and `memchr` runs plus `memchr#215` and `rust-smallvec#407` PR-diff runs; dogfood found and fixed import/declaration false positives, `cfg(target_feature)` false positives, capped repo scan timeout behavior, missing owner-contract inheritance for operation cards, and comment-derived owner false positives | Experimental | More crates, more real PR diffs, uncapped/sampled runs, and human review are still needed before calibration claims |
 
 ## Current Gaps
 
diff --git a/docs/status/SUPPORT_TIERS.md b/docs/status/SUPPORT_TIERS.md
index 35361bd4c..d92d93dc1 100644
--- a/docs/status/SUPPORT_TIERS.md
+++ b/docs/status/SUPPORT_TIERS.md
@@ -12,7 +12,7 @@ All tiers describe static review evidence. None means memory-safety proof.
 | Raw pointer card slice | experimental | cards | `raw_pointer_alignment`, `raw_pointer_deref`, `raw_pointer_read_unaligned`, `raw_pointer_write_assignment`, `split_raw_pointer_read_call`, `split_unsafe_block`, and safe-reference negative fixtures | source-level review evidence only |
 | Core operation smoke slice | experimental | cards | `maybeuninit_assume_init`, `vec_set_len`, `transmute_invalid_value`, `get_unchecked_mut_bounds`, and `pin_new_unchecked` fixture goldens | curated fixtures, not broad semantic proof |
 | Fixture calibration manifest | experimental | `fixtures/calibration.toml` / `cargo xtask check-calibration` | manifest covers positive, negative, and false-positive-control core fixture claims, including import/declaration and `cfg(target_feature)` false-positive controls, and validates expected card counts, classes, operation families, and hazards against goldens | proof index only; not real-world calibration or support-tier promotion evidence by itself |
-| Real-crate dogfood measurement | experimental | dogfood handoff | `2026-05-18-real-crate-dogfood-v0.6.md` records top-50 capped `rust-smallvec`, `arrayvec`, and `memchr` repo runs plus a `memchr#215` PR-diff run; it also records the import/declaration, `cfg(target_feature)`, capped-scan, and owner-contract inheritance fixes those runs exposed | three capped snapshots and one PR diff only; no calibrated rates, no full audit, no uncapped performance claim, and no support-tier promotion |
+| Real-crate dogfood measurement | experimental | dogfood handoff | `2026-05-18-real-crate-dogfood-v0.6.md` records top-50 capped `rust-smallvec`, `arrayvec`, and `memchr` repo runs plus `memchr#215` and `rust-smallvec#407` PR-diff runs; it also records the import/declaration, `cfg(target_feature)`, capped-scan, owner-contract inheritance, and comment-aware owner inference fixes those runs exposed | three capped snapshots and two PR diffs only; no calibrated rates, no full audit, no uncapped performance claim, and no support-tier promotion |
 | Contract evidence mining | experimental | cards | public unsafe fn/trait fixtures and private helper `SAFETY:` fixture | comment quality is heuristic |
 | Guard evidence mining | experimental | cards | raw-pointer alignment and comment-not-guard fixtures prove bounds evidence does not discharge alignment | obligation-specific patterns are still sparse |
 | Witness routing | experimental | cards | route-table tests plus raw pointer, FFI, unsafe impl Send, Pin, and invalid-value fixture routes | route recommendation only; receipt import is a separate surface and does not execute witnesses |