docs: refresh the benchmark baseline for the :string fast path; mark v0.5 item 9 done

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DPAgXWnvgpKcUw71GHStXC

Author: pegeaether

ROADMAP.md

@@ -53,8 +53,13 @@ docs/design/v05-plan.md (three-proposal panel, two verifying judges).
    generated code, `make check-versions` (fails on a skewed site),
    docs/releasing.md, and `SBCL / Linux aarch64 (best-effort)` — green on
    its first run.
-9. **`:string` ASCII fast path** — a typed ASCII loop before babel (both
-   judges measured ~2.5×); zero-copy `:string` stays refused.
+9. ✅ **`:string` ASCII fast path** — a typed check-and-store loop in
+   both directions, babel from the first char/byte ≥ 128, a peek before
+   any allocation so non-ASCII text pays nothing extra: 64 KiB ASCII
+   1015 → 318 µs (3.2×), non-ASCII unchanged. Differentially tested
+   against babel on all three hosts (no counterexample in ~12k checks)
+   and swept adversarially through echo; zero-copy `:string` stays
+   refused.
 10. **Miri over the generated shims** — optional; cut first.
 11. **User-facing docs claim audit, last** — README still says
     "Scope (v0.1)"; every capability claim gets a citation or is deleted.

docs/benchmarks.md

@@ -15,7 +15,7 @@ compilation policy; nothing is tuned. Numbers vary by machine and are
 nanoseconds, byte and vector transfers scaling linearly at memcpy speed,
 a borrowed callback a small multiple of a call.
 
-## Baseline — 2026-09-02, commit 5241783
+## Baseline — 2026-09-03, commit 43f3c9c
 
 | | |
 |---|---|
@@ -27,27 +27,28 @@ a borrowed callback a small multiple of a call.
 Raw output of `make bench`:
 
 ```
-scalar call (add i64 i64)                      20.0 ns/call
+scalar call (add i64 i64)                      25.0 ns/call
 scalar call, no args                           25.0 ns/call
-handle method (gate + call)                   240.0 ns/call
+handle method (gate + call)                   275.0 ns/call
 handle create + free                          650.0 ns/call
-string round trip (8 B ASCII)                 950.0 ns/call
-string round trip (1024 B ASCII)            22701.5 ns/call
-string round trip (65536 B ASCII)         1015066.3 ns/call
-string round trip (1 KiB non-ASCII)         39452.6 ns/call
+string round trip (8 B ASCII)                 600.0 ns/call
+string round trip (1024 B ASCII)            11100.7 ns/call
+string round trip (65536 B ASCII)          317620.7 ns/call
+string round trip (1 KiB non-ASCII)         38802.6 ns/call
+rx count over 1 MiB (&str in, 1 match)    2305160.0 ns/call
 bytes in (sum, 8 B)                           200.0 ns/call
 bytes round trip (rev, 8 B)                   500.1 ns/call
-bytes in (sum, 1024 B)                        500.1 ns/call
-bytes round trip (rev, 1024 B)               1200.0 ns/call
-bytes in (sum, 65536 B)                     21251.3 ns/call
-bytes round trip (rev, 65536 B)             39952.6 ns/call
-bytes in (sum, 1048576 B)                  323521.0 ns/call
-bytes round trip (rev, 1048576 B)          731048.0 ns/call
-vec i64 round trip (8 elts)                   650.0 ns/call
-vec i64 round trip (1024 elts)               6600.4 ns/call
-vec i64 round trip (65536 elts)            192512.5 ns/call
-borrowed callback (100 invocations)         48003.0 ns/call
-stored callback (same thread)                 180.0 ns/call
+bytes in (sum, 1024 B)                        550.0 ns/call
+bytes round trip (rev, 1024 B)               1100.0 ns/call
+bytes in (sum, 65536 B)                     22451.4 ns/call
+bytes round trip (rev, 65536 B)             44102.9 ns/call
+bytes in (sum, 1048576 B)                  319521.5 ns/call
+bytes round trip (rev, 1048576 B)          702045.0 ns/call
+vec i64 round trip (8 elts)                   600.0 ns/call
+vec i64 round trip (1024 elts)              10550.7 ns/call
+vec i64 round trip (65536 elts)            191012.5 ns/call
+borrowed callback (100 invocations)         39002.5 ns/call
+stored callback (same thread)                 170.0 ns/call
 ```
 
 ## Reading it
@@ -68,12 +69,18 @@ stored callback (same thread)                 180.0 ns/call
   (0.32 ms and 0.19 ms) with the element-wise marshaller they replaced
   (7.67 ms and 57.3 ms, measured the same way on this host before the
   change — see the v0.3 groundwork commit).
-- **Strings are the slow path**, and deliberately so for now: a
-  `:string` round trip goes through UTF-8 encode/decode on the Lisp
-  side, so 64 KiB costs ~1 ms where the same bytes as `:bytes` cost 40 µs.
-  Zero-copy `:string` is refused-with-cause in the v0.3 plan (bodies are
-  octets; no demand case); pass large payloads as `:bytes` and decode
-  what you need.
+- **Strings: ASCII takes a typed loop, the rest goes through babel.**
+  A `:string` crosses as UTF-8 with a copy in each direction (BOUNDARY
+  §4; zero-copy `:string` is refused-with-cause in the v0.3 plan). Since
+  v0.5 the loader checks-and-stores ASCII text in a typed loop and hands
+  the bytes to babel at the first char/byte ≥ 128, so a 64 KiB ASCII
+  round trip costs ~0.32 ms where the 2026-09-02 baseline (in git
+  history) measured ~1 ms; the same bytes as `:bytes` cost 40 µs, so
+  octet payloads still belong in `:bytes`. Non-ASCII text costs what it
+  did (the 1 KiB row). The rx row is a real `&str` consumer: 1 MiB into
+  `regex-count` with the digits placed once at the end, so the regex side
+  is a memchr sweep and the row is the boundary — ~2.3 ms, in line with
+  16× the 64 KiB row.
 - **Callbacks**: a borrowed callback invocation is ~480 ns (100 per
   export call); a stored callback from the same thread is ~180 ns. The
   cross-thread stored path adds foreign-thread adoption, not measured