Record the QUIC Linux leg green, and a README path that never worked

All ten scenarios pass under asserts and all four required under release, on
Yggdrasil, in the committed debian:13 image. Worst overshoot anywhere was 4 ms
against a 400 ms tolerance. F1 is closed on Linux: the build that ships now
runs the timeout rule.

Running the README's own worked example is what found the second thing. It
cannot work: the test binary carries an $ORIGIN rpath and nothing puts
libmsquic beside it for a test build. That example was written in the same
commit that deleted the harness, so the working path was removed and the
broken one published together.

Two things stay open. Windows still owes its release configuration for the
same reason Linux did. And F5's stress wants sixteen CPU burners on the
platform where burners are now forbidden, which is the operator's call rather
than mine to route around.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

docs/typescript-quic-realtime-cut.md

@@ -836,9 +836,47 @@ argument to the binary. **The release configuration is therefore owed again as
 a plain build, not as a harness target**, and R1's substance — that the
 release macro is unexercised — is a Soul hand probe now.
 
-**Outstanding for Cut 3: the Linux leg.** Every scenario, asserts and release,
-in the committed `debian:13` dev image, run on Yggdrasil. Windows is the
-workstation's only remaining QUIC work.
+**The Linux leg ran on Yggdrasil, 2026-09-22** (Sonnet), against
+`hands/cultmath-erf` HEAD, in an image built from the committed
+`scripts/quic-native-linux-dev.Dockerfile` (pinned `debian:13`, tagged by
+content hash `cultlib-quic-native-dev:13d7fbd109fc`). No repo change was
+needed and none was made.
+
+- **Asserts build, all ten scenarios pass**, 20 rounds each: `closerace`,
+  `holdclose`, `holdtimeout`, `polltimeout`, `pollbusy`, `pollhammer`,
+  `zerotimeout`, `payloadfit`, `latecall`, `waitseam` (11 probes, 40 repeated
+  polls on one runtime).
+- **Release build, asserts off, all four required scenarios pass**:
+  `polltimeout`, `pollbusy`, `pollhammer`, `zerotimeout`.
+- Worst overshoot anywhere was 4 ms, against `kGenerousLateToleranceMs` of
+  400 ms. **F1 is closed on Linux**: the build that ships now runs the timeout
+  rule.
+
+**A defect in the README's own worked example, found by running it.** The
+example builds the test binary and runs it directly, and every scenario dies
+with `libmsquic.so.2: cannot open shared object file`. Both the bridge and the
+test executable carry `INSTALL_RPATH "$ORIGIN"` and expect the MsQuic runtime
+beside them in `bin/`, but nothing puts it there for a test build — only
+`build-quic-native.sh`'s packaged output directory gets that copy. Hands
+worked around it with a `cp` after each build and correctly did not treat it
+as a bridge defect. **This is the documented path, written in `796433f` when
+the harness was deleted, and it has never worked.** A documented path that
+fails on first use is worse than no documentation: it was published in the
+same commit that removed the working one.
+
+**Still outstanding for Cut 3:**
+
+- The **Windows release configuration**. It lived inside the harness and left
+  with it, exactly as Linux's did; Linux has now been re-run without the
+  harness and Windows has not.
+- **F5's stress under load cannot be run on Windows at all.** It wants five
+  rounds with sixteen CPU burners, and those burners are what froze the
+  operator's workstation on 2026-09-22. `closerace` is Windows-sensitive (it
+  went red under load there in four of five rounds), so the check matters most
+  on the platform where it is now forbidden. **This is an operator fork, not
+  something to route around:** either Windows stress finds a host that is not
+  Starfire, or the check is accepted as unproven on Windows and recorded that
+  way.
 
 **Cut 1 Soul findings, 2026-09-16.** Held: every test count, both negative
 greps, all twelve mutations rerun and killed, the control catching a