A #\' near-miss errors in value position and back-points the quote it strands elsewhere

The raw-string delimiter is a run of # immediately followed by ' — #', ##', ###' — and a #-run not followed by ' opens a comment to end of line. The mistype #\' therefore falls in a blind spot: the backslash breaks the delimiter, the run opens a comment, and the rest of the line is silently swallowed. The failure then surfaces far from its cause — a later lone ' raises [L0001] unterminated, anchored at the close, while the stray backslash at the open goes unmentioned. This is the trap documented in the Strings section of exarch/data/ral.md (echo #\'…\'# > f “writes nothing and exits 0”); it costs a reader many turns chasing the wrong line.

The fix is diagnostic, not grammatical: #' still opens a raw string, a bare #-run still opens a comment, and #\' is still not an opener (raw strings escape nothing — there is no escape to honour). What changes is that the lexer no longer lets the near-miss pass in silence.

Position is the signal of intent, so the diagnostic splits on it

A #-run reaches the lexer by one of two paths, and the path discloses what the author meant:

• Value position — the run is dispatched by next_token’s # arm, reached after inline whitespace or at start of input (echo #\'…, the canonical reproduction). Here a raw string is the natural reading, so a near-miss is a hard lex error, [L0004], anchored at the #\' span and naming the real delimiter.
• Line-leading position — the run is consumed by scan_separator’s comment skip, immediately after a statement separator. Here a # reads naturally as a comment, so the run stays a comment — but its span is recorded, and if a quote later dangles, [L0001]‘s secondary label points back at it as the likely cause.

The asymmetry is deliberate. Rejecting a line-leading #\' would change the grammar of comments; leaving a value-position #\' silent is exactly the trap. Each path gets the treatment its position warrants.

The mechanism: one new error kind, one near-miss memory

• LexErrorKind::RawDelimiterEscaped { hashes, quote, span } → [L0004], joining the structured lexer codes [L0001]–[L0003]. Its message is run-length aware (raw-string delimiter is `##'`, not `##\'` for a two-hash run); the primary label reads remove the backslash, with the rule in the hint.
• UnterminatedString gains hash_escape: Option<Span>, filled from a lexer field that records the most recent line-leading near-miss. The field is cleared on every successful string close, so a stale near-miss cannot haunt an unrelated later dangling quote — the heuristic fires only when no complete string intervened between the near-miss and the open quote.
• Rendering (diagnostics) repurposes UnterminatedString’s single secondary-label slot: where today it points at EOF (expected closing ' here), a recorded near-miss redirects it to the #\' span (this # opened a comment, not a raw string), plus an explanatory hint.

Both detections share one predicate — a #-run of length n followed by \' or \" — so value-position rejection and line-leading recording cannot drift apart.

Consequences

• echo #\'…\'# > f changes from a silent no-op (exit 0) to a hard [L0004] error. This is the one observable behaviour change, confined to value position; the Strings section of exarch/data/ral.md is rewritten to describe the rejection rather than the silent swallow.
• A line-leading #\' comment is unchanged unless it strands a later quote; then the [L0001] it provokes names it.
• needs_continuation does not match RawDelimiterEscaped, so a REPL line echo #\'… reports the error at once rather than waiting for a continuation that will never arrive.

Reading of “do not change the grammar” (load-bearing — needs confirmation)

The whole shape above rests on reading “#\' stays a comment” as “do not reclassify #\' as a raw-string opener”: the #' rule is untouched, and the lexer merely refuses the suspicious comment with a diagnostic. Under that reading the value-position case is a hard error. The alternative reading — the byte sequence must still tokenise as a comment, only the surrounding diagnostics improve — forbids the hard error and collapses the fix to the line-leading back-pointer alone. This is the one decision the design cannot make for itself.

Open questions

• Dedicated variant vs. Other. A free-form Other(msg) would carry the message without a stable code or a tailored label. The dedicated RawDelimiterEscaped matches the module’s structured-arm design and gives the doc an [L0004] to cite. Proposed: dedicated.
• \" lumped with \'. #\" is diagnosed by the same arm and named against #'; “remove the backslash” leaves #", still not a delimiter, so the advice is imperfect for the double-quote case. Proposed: keep them together — the message still names the right delimiter.
• Clearing on close. Keeping the near-miss memory only until the next successful string close is what stops false positives across unrelated quotes. Proposed: keep.
• One secondary slot vs. many. Pointing at the near-miss costs the EOF pointer, because render_ariadne carries a single secondary label. Widening it to a list is a larger change to a shared path. Proposed: repurpose the one slot.

See also syntax, diagnostics, surface-syntax; the lexer rules at core/src/syntax/lexer.rs (the module docstring, the # dispatch in next_token, and scan_separator).