breaker: no protocol in kernel; the transport states its fault table, PIECE_END=128 rides the same framer lineage (row 66)

[MattJackson]

Row 66, "breaker: no protocol in kernel". The kernel's breaker no longer classifies from status bands or numberings. It reads the fault the transport states (and, since #489, the plane's), and adds one fact of its own: that no answer came.

Spec: docs/design/BUSBAR-1.6.0.md Part 0 (l.115-120: a transport may name only itself; l.3962: no transport names another), Part 2 #36 (busbar-kernel-breaker: trip, cooldown and fast-fail per pool and destination; the kernel names no protocol), Part 3 (the route pump: the breaker is the kernel walk's, and the breaker service classifies).

What changes

• Transport fault ABI (busbar-contract):
  • FramePiece::fault takes the spare byte, so the piece size is unchanged.
  • FaultRow and TransportTail::fault_rows are appended to the tail. A tail that predates them reads its fault table as absent (TRANSPORT_TAIL_FROZEN).
  • check_fault_rows and check_fault_cover: a claim that states status rows must state fault rows, or the load is refused. Both have RED arms.
  • The layout golden and the C header are regenerated.
• kernel-breaker:
  • classify.rs is Reading{NoAnswer, Answered(Option<WireFault>)} -> Disposition. HTTP_OVERLOADED/529, the gRPC status table, the HTTP bands, the context-length gate and the warn-once diagnostics are gone.
  • port.rs UpstreamStatus is {reading, retry_after}.
  • The crate names no plane, no header and no vendor. The kernel-breaker x plane row in kind-isolation is struck.
  • normalize.rs stays, marked NOT A BREAKER INPUT: only the retiring engines import it.
• Walk and root: the far-end walk records from the reading. The root adapter reads Answered(fault), and reads NoAnswer only when there is neither a class nor a fault.
• PIECE_END=128 (the PIECE-END lane's p2-piece-end, c52ba30, merged in whole): the stream's end is said, never an empty piece. check_framer rules, sdk_hyper Piece::end, connector Got.end.
• Transports: one lineage carries the fault table and PIECE_END. Each repo's branch is p2-breaker-neutral-ts-pin = dev + p2-piece-end-b, repinned to busbar 2add33a:
  • tcp b3047ce874: transport stage B: carriers over the host's I/O (tcp, stdio), the connector drives every connection through a carrier #502's carrier, plus its tail's empty fault table
  • ws 2fd38617db
  • http 0ff0e022e8
  • grpc 1303273a73: the door's one-stream framings are renamed streams -> calls, because a transport names only itself (kind-isolation vocab)
• The neutral frame door fixture states its fault table (claim 0, 1..=9 FAULT_CALLER), so it trips nothing, exactly as before.

Behaviour

1.5.5-identical. The http framer's fault table maps exactly onto the old bands:

• 401/403: hard
• 408/429 and 5xx (529 included): transient
• other 4xx: caller
• off-table: none, which reads as caller
• no answer: network, which is transient

Proof (Latchkey, large)

• Tests, unfiltered: kernel 3001, kernel-breaker 27, kernel-egress 128, contract 1250, connector 250, plugin-loader 701, busbar 845. The tcp/http/ws/grpc plugin cdylibs are built for the both-ways tests.
• Lint: clippy -D warnings on the workspace with all targets, and fmt.
• Kind-isolation: equal to predev. deps has the same 4 findings and test-deps the same 5, compared finding by finding; vocab has 0.
• Oracle, strict against golden/1.5.5, 54 cells (route.529, route.failover, cooldown, crosscut.traps, billing, ledger, teller, concurrency):
  • Replay rows are identical between this branch and its base, recorded on the same runner.
  • The 19 FAIL rows are predev's own (table in ORACLE-BURN's hands; P6 ORACLE-STRICT: strict-replay reds cleared or routed (auth_modules, register under signed rulings, engine pin 2119a67a6f, bin/oracle Q128, scrape-reload cell, 1.6 plugin source pins) #493 clears them).
  • The 21 failover/529/breaker cells did not move.
• Autobahn ws suite, strict: pass.

[MattJackson]

Marked draft: every change here is carried by #787 (verified: head/patches contained). Draft removes it from ModernCI's blocked-by set (ModernCI #292/#308) so #787 is judged on its own. It shows as merged once #787 lands.

[MattJackson]

Closed: every commit of this PR is patch-identical to content that landed on predev in the wave-1 integration (predev 179abf2, lane-integration-w1b). Nothing remains to merge.