Record the M3.13e delivery and the next slice

[kaikisegfault]

Documentation only. docs/project/current-state.md, no source, build, workflow,
or vector change, so this is the focused metadata verification path.

What it records

• M3.13e is delivered: decode_genesis and protocol-application-v7 are on
  main across cf1d28c..8ffc2bd, and a version-seven node is a process that
  can be started, connected to, restarted, and shut down.
• Why the decoder is the encoder's inverse and checks itself, what the round
  trip catches that a restatement would not, and the field-order trap it exists
  for: total_supply is encoded before fixed_transfer_fee while the struct
  declares them the other way round.
• The failure the hosted matrix caught that no local check could. A target
  declared but never added to PROTOCOL_STACK_TARGETS builds at the compiler's
  default standard, and the scratch harness passes -std=c++20 on every
  invocation, so it compiles clean locally and fails in all four hosted jobs with
  errors pointing at headers that have not changed in months. The guard added
  with the fix makes it non-repeatable, and the handoff records that adding an
  executable now means four edits rather than three.
• Two probes that passed without being gaps, and why each was right to:
  one changed an allocation bound rather than a rule, and one aimed at a rule
  the application suite owns — which was re-run to confirm the coverage rather
  than assumed.

The next slice

M3.13f, the version-seven ABCI adapter — the last structural piece between
the stack that exists and a network that runs it. The entry records that
adapter/cometbft already speaks the same frame format, that only three things
differ (the block identifier in the finalized-block response, the app state
string, and the wider result-code table), and that the slice's one real design
question is not encoding at all: what the adapter does about the replay
handshake, since ApplicationV7 refuses terminally a finalize_block at any
height that is not the next one, including one it has already committed, which is
exactly what CometBFT does to an application behind its engine.

A note on dates

The deliveries are dated 2026-08-31 because that is what their CI evidence shows;
the handoff itself is written after local midnight, so Last updated is
2026-09-01.

Verification

python3 -B tools/verify_metadata.py — 5 repository skills and 265 internal
Markdown links validated. git diff --check clean. The hosted focused metadata
path is the gate for this exact commit.