The missing version field, named; positioning and walk-backs made standing docs

Authored by the parallel "MOQ over time since 1999" session, reviewed and
committed here after verifying its central claim against the code: Wire.VERSION
(0x54530000) only tags the build - nothing writes a version into any packet, so
today a wire-format skew between builds is an undiagnosable decode failure
rather than a named mismatch. SPEC had listed "version negotiation + greased
versions" in the Keep column for all of v0; it was never built.

- docs/TODO.md gains section 11: what would settle it (a version field, a defined
  mismatch response, and a two-build test asserting the receiver NAMES the
  mismatch), why waiting compounds (every pre-field wire change is one no
  deployed peer can detect, and the independent decoder of section 7 will encode
  whatever v0 is on the day it is written), and that it is blocked on nothing.
- docs/SPEC.md moves the row to a new "Claimed, not implemented" column instead
  of deleting it - the gap between claim and code is the interesting part - and
  fixes the :288 note that would have gone stale with the move.
- CLAUDE.md gains "Why this exists" (the no-Google rule is the experiment, not
  licensing hygiene; RLNC is the load-bearing bet; forced convergence back onto
  an RFC is data) and a standing walk-backs list, seeded with the Retry
  reversal and this inverse case - never rejected, never built, caught by
  pinning golden wire vectors rather than by reading the ledger.

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

Author: 3x3xX3N0N

CLAUDE.md

@@ -10,6 +10,36 @@ Homa, RLNC literature, CUBIC/HyStart++, and IETF specifications implemented from
 records what was deliberately borrowed and what was rejected; `NOTICE` records provenance per mechanism. Keep
 that table current when adding anything.
 
+## Why this exists (read before touching the constraint above)
+
+The no-Google rule is not licensing hygiene, it is the experiment. Tessera re-runs the choice the transport
+world made in 2000, when SCTP failed to deploy because middleboxes dropped an unknown IP protocol number, and
+QUIC's answer a decade later was to hide a new transport inside UDP. Tessera takes that answer and declines
+the rest of the package: it sits at QUIC's layer rather than above it, where MoQ and WebTransport sit.
+
+Two things follow, and they are the reason to keep the provenance table honest:
+
+- **The load-bearing bet is RLNC.** Every deployed transport on both the web and media lineages recovers loss
+  by asking again — at minimum one RTT per loss — and nothing has revisited that since 1999. Coded repair is
+  the only mechanism here that could not be obtained by configuring an existing QUIC stack. If the flat-tail
+  claim survives real radios, that is the result. Everything else is a defensible variation.
+- **Forced convergence is data, not defeat.** Implementing RFC 9000/9001/9002 from the documents makes this an
+  independent reading of them. Every place the reimplementation *had* to come back to the spec says something
+  about which parts of QUIC are essential rather than incidental. Those are findings; write them down.
+
+## Walk-backs — mechanisms rejected, then readopted
+
+A standing list, because this class of event recurs and is worth more than the individual fixes. When a
+"leave behind" decision reverses, add it here as well as in the `docs/SPEC.md` ledger.
+
+- **Retry / address validation** (v0.7). Originally rejected as QUIC ceremony. Wrong: the amplification limit
+  and ticket binding bound *reflected bytes*, not *CPU*, and the ML-KEM operation ran before anything had
+  authenticated the sender — a cost RFC 9000 could not have anticipated but that its mechanism happens to
+  cover. Retry now ships, gated on pressure so 0-RTT survives the common case. `docs/SPEC.md`, the struck row.
+- **Version negotiation** (open, `docs/TODO.md` §11). The inverse case: never rejected, listed as kept for the
+  whole of v0, and never built. Caught by pinning golden wire vectors, not by reading the ledger — which is
+  the argument for pinning things.
+
 ## Layout
 
 | Module | Contents |

docs/SPEC.md

@@ -127,7 +127,6 @@ Actually open:
 | Amplification limit (3×) before path validation | Only sane anti-reflection design |
 | ACK ranges + ECN counts + ack-delay field | Cheap, precise loss/OWD signal |
 | Transport-parameter TLV with grease | Extensibility that survives middleboxes |
-| Version negotiation + greased versions | Ossification insurance |
 | Header protection (encrypt PN) | Prevents middlebox PN inference; cheap |
 | Key update via key-phase bit | Re-key without handshake |
 | DPLPMTUD | Real MTU, fewer packets |
@@ -145,6 +144,10 @@ Actually open:
 | HTTP/3-shaped priorities | App-layer concern |
 | Multipath as extension | Multipath is native: per-path PN space, cross-path repair |
 
+| Claimed, not implemented | State |
+|---|---|
+| Version negotiation + greased versions | The intent was ossification insurance; v0 packets carry no version field, and `Wire.VERSION` only tags the build. Listed here rather than under **Keep** because the ledger a new reader trusts first should not describe a mechanism that is absent. Open item: `docs/TODO.md` §11. |
+
 ## v0.3 (2026-08-22) — eight parallel modules merged, 74 tests
 | Module | File | Status |
 |---|---|---|
@@ -285,9 +288,10 @@ described something the code does not do. The code is authoritative; the spec wa
   1-byte PN is parseable but never emitted. The short header is 7 bytes typical, not 6.
 - **`0x81` is Padding**, carrying 2–257 bytes, used for PMTUD probes. Previously the `0x80+` range was described
   only as "extension/grease" with no assignments listed.
-- **There is no version field on the wire.** "Version negotiation + greased versions" is listed among the
-  mechanisms kept from QUIC, but v0 packets carry no version; `Wire.VERSION` only tags the build. It remains a
-  design intention, not an implemented mechanism, and is listed under open items instead.
+- **There is no version field on the wire.** "Version negotiation + greased versions" used to sit in the
+  **Keep** column, but v0 packets carry no version; `Wire.VERSION` only tags the build. It is a design
+  intention, not an implemented mechanism, and now sits under **Claimed, not implemented** above, with the
+  open item at `docs/TODO.md` §11.
 
 ### Closed: native dual-stack on Windows
 

docs/TODO.md

@@ -155,6 +155,26 @@ Ten Vultr nodes (`sao scl jnb syd blr sgp mex icn ewr del`) are up and billing a
 `python bench/mesh/mesh.py destroy` is the off switch. `bench/mesh/shape.py clear all` before destroying if any
 node is still shaped — and verify with `show`, because `clear` has lied once already.
 
+## 11. No version field on the wire — the upgrade path does not exist
+
+`docs/SPEC.md` listed "version negotiation + greased versions" among the mechanisms kept from QUIC for the
+whole of v0. It was never built: v0 packets carry no version, and `Wire.VERSION = 0x54530000` only tags the
+build. The row has been moved to a **Claimed, not implemented** column rather than deleted, because the gap
+between the two is the interesting part.
+
+This is not cosmetic. `README.md` says to expect the wire format to change, and `WireVectorsTest` notes that a
+peer built from an older commit no longer interoperates — so today a version skew is an undiagnosable decode
+failure rather than a clean mismatch. Ossification insurance is the one lesson the whole design inherits from
+SCTP's failure to deploy, and it is the lesson currently not applied.
+
+- **What would settle it:** a version field in the long header, a defined mismatch response, and a test that
+  runs two builds with different versions against each other and asserts the *receiver names the mismatch*
+  rather than failing to parse. Greasing is a second step and not required to close this.
+- **Cost of waiting:** it compounds. Every wire change made before there is a version field is a change no
+  deployed peer can detect, and the independent-decoder work in §7 will encode whatever v0 happens to be on
+  the day it is written.
+- **Not blocked on anything.** No hardware, no nodes, no measurement — this one is only unwritten.
+
 ---
 
 ## Closed, with the argument worth keeping