[tensorlink-dev]

Owner-approved 2026-11-13 04:30 UTC ("Build this"). No runtime behaviour change, no consensus field touched, no restart needed.

Why

Decided on-chain votes (DEC-CA-0045 stake-weighted activation) are meant to be TYPED into the release chain.toml once applied — behaviour-identical, and it frees the feature's segment of the 128-byte readiness note. Until tonight that typing happened only on the operator's deployed file. main lagged for weeks; an external validator that pulled main overflowed its note and refused to start (OPSLOG 2026-11-13 00:13; symptom fixed by #457). Nothing connected "what the fleet decided" to "what main ships". This PR does.

What

1. decided_votes.toml (beside chain.toml) — the one place in git recording what the fleet decided: each vote's feature name, lock block, activation block, source, and the key that types it; plus the four worker-image pins the network trains on. Rows can be type_it = false (HELD) with a reason: rolling-era-king (nodes resolved it differently — ours 9133200→9136800, one external 9374400→9378000 — so main must not type either), and prelaunch-95-r2 (decided 9486000→9486900 everywhere seen; typing it is behaviour-identical and frees another segment, but it is a release chain.toml edit awaiting the owner's word — next one-line PR).
2. cascade/shared/decided.py — drift_report(cfg, record): TYPE (decided, applied, not typed → prints the exact [section] / key = block lines), MISMATCH, PIN_DRIFT, STALE (a row whose feature name the config no longer computes = changed vote params = a fresh vote), PENDING (not yet applied as of the record's block — never asked early), HELD. state_report(records, record): a node's activation_*_state.json lock-ins vs the record → UNRECORDED (prints the [[vote]] row to add), DISAGREE, UNKNOWN (retired feature; our box still holds a promotion-carry record).
3. cascade-audit decided [--config F] [--decided F] [--state-dir D] [--as-of-block N] [--json] — runs both reports, exit 1 on drift. Reads only.
4. CI guard — tests/unit/test_decided_votes.py::test_repo_chain_toml_and_record_are_in_sync runs drift_report on the repo's own files: a vote that resolves and is not typed (or a pin moved on one side) turns CI red on the next PR, whatever it is about. Plus unit tests for every finding kind and the CLI.
5. Lock-in log line — the resolver (resolve_activation) now logs DECIDED: <feature> applies at block N — once that block has passed, type \key = N` under [section] in the release chain.toml and record it in decided_votes.toml (PR to main)` at every LOCKED IN. One lazy import; nothing else changes.
6. docs/AUDIT.md section.

Verified on the box (python -P, config + stdlib only)

• cascade-audit decided on the repo: in sync (4 OK votes, 2 HELD, 4 OK pins), exit 0.
• On the deployed file with our state dir: reports exactly the known gap — [TYPE] v2b-69630f92 → [v2] image_decoupled_from_block = 9493200 (to add at the next quiet validator restart), the HELD prelaunch row, [UNKNOWN] promotion-carry (retired feature's state file). Pins OK. Exit 1.
• ruff clean.

Not in this PR (owner decisions)

• Typing [v2] prelaunch_burn_from_block = 9486900 in main (HELD row above).
• Publishing the netconfig overlay (validators would get decided values within 5 min without pulling main) — testnet first, validator restart.

🤖 Generated with Claude Code

https://claude.ai/code/session_01BMT4D7eZUKUF2gocYLVFmj

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
cascade	Ready	Preview	Nov 13, 2026 4:19am UTC