fix(observability): bound the MPC metrics hot path; refuse mock-crypto builds on real networks

[omersadika]

What

Item 7 of the 1.2.0 pre-tag checklist (release review) — the cheap fixes worth taking before the tag.

1. MPC metrics hot path (hand-verified findings)

refresh_observability_metrics ran on every ~20ms service iteration and re-emitted ~11 per-seq gauge series (plus a per-committee-member matrix) for every session tracked this epoch — terminal sessions stay in self.sessions until epoch end (zero remove/retain), so per-tick work grew with the epoch's cumulative session count, on the same loop that advances MPC sessions. The in-code "bounded by max_active_sessions_buffer" claim was false.

• Refresh rate-limited to 1/s (OBSERVABILITY_REFRESH_INTERVAL).
• Terminal (Completed/Failed) sessions get exactly one final emission recording the terminal state, then are skipped; re-armed if a session ever leaves terminal state. Their seqs stay in current_seqs so the existing stale-zeroing block doesn't wipe the terminal record.
• The sessions_with_self_output_no_quorum counter moved into the cheap unconditional scan — semantics unchanged.

2. Checkpoint per-session gauge label leak

ika_dwallet_checkpoint_user_session_written_at_seq accumulated one labeled series per user session for the process lifetime. The builder is per-epoch, so its constructor now reset()s the gauge — label set bounded to the current epoch.

3. unsafe-mock tripwire

A binary compiled with dwallet-mpc-unsafe-mock now refuses to boot against ika mainnet/testnet (chain derived from the deployed ika system object, exactly like #1810's minimum-CPU gate; predicate unit-tested in the same style) and logs a grep-able insecure_mock_crypto=true error banner on any other network. CLAUDE.md documents the verified footgun that workspace-scope cargo test silently mocks ika-core's crypto via ika-test-cluster's self-dev-dependency feature unification (verified both directions with cargo tree: dev-graph resolution enables it; plain ika-node builds do not).

4. Deployed-bytes decode tool (G1 evidence residual)

New #[ignore]d decode_deployed_network_key_bytes (sharing the deployed-envs table with the existing NetworkKeyId tool): fetches the real mainnet/testnet network-key bytes and decodes them on the exact types the v3 backward-compatible reconfiguration arm consumes — V1 anchor inner → class-groups DKG output (+ byte-identical re-encode check), V2 reconfiguration output → bwd-compat PublicOutput. This is the real-chain-bytes check the synthesized-anchor regression tests can't provide; run results will be posted on this PR.

Verification

• cargo clippy -p ika-core --lib --tests clean; cargo fmt --all.
• runtime::tests 4/4 green (new mock_crypto_forbidden_exactly_on_real_networks included).
• Behavior notes: rate limit and once-per-terminal emission change gauge freshness (1s) only; series values and meanings are unchanged. Registry cardinality within an epoch is unchanged (per-seq series still exist); steady-state per-tick work drops from O(epoch sessions × 50/s) to O(in-flight sessions × 1/s).

🤖 Generated with Claude Code

[omersadika]

Deployed-bytes decode evidence (G1 residual: CLOSED)

Ran decode_deployed_network_key_bytes (release build, inkrypto pin 1d8b5a9, 2026-07-12) against both live chains:

CONNECT testnet: https://fullnode.testnet.sui.io:443 failed: Request rejected `404`
CONNECT testnet: using https://sui-testnet-rpc.publicnode.com
DECODE testnet: key 0xe7c79a60931299e110297554fc02e0a0e095e96778775092c97f07a1bd1337cc: V1 anchor OK (175302 bytes, re-encode byte-identical)
DECODE testnet: key 0xe7c79a60931299e110297554fc02e0a0e095e96778775092c97f07a1bd1337cc: V2 reconfiguration output OK (151034 bytes)
CONNECT mainnet: using https://fullnode.mainnet.sui.io:443
DECODE mainnet: key 0x0a9c0b88a5c729378bce1a98a8c285f1dde26f89e53d07164dba8a059dc15587: V1 anchor OK (180458 bytes, re-encode byte-identical)
DECODE mainnet: key 0x0a9c0b88a5c729378bce1a98a8c285f1dde26f89e53d07164dba8a059dc15587: V2 reconfiguration output OK (197504 bytes)
test ... ok (finished in 40.15s)

What this proves, on the actual on-chain bytes (not synthesized test vectors):

1. Both deployed anchors are V1-tagged — the premise of the G1 fix (fix(dwallet-mpc): handle the deployed V1 network-DKG anchor on the v3 reconfiguration path (G1) #1798/feat(dwallet-mpc): support the deployed V1 network-DKG anchor on the main (v4) reconfiguration path + one-time V1->V3 anchor migration #1799) holds on both networks.
2. The exact decode the v3 backward-compatible reconfiguration arm performs (V1 inner → class_groups::dkg::PublicOutput) succeeds on the real bytes, and re-encodes byte-identically — no bcs layout drift across the crypto-source migration and both inkrypto bumps (0572d6c → c407ad4 → 1d8b5a9).
3. The prior V2 reconfiguration outputs decode as the bwd-compat PublicOutput — the second input the arm consumes on every reconfiguration of the deployed keys.

This closes the last G1 residual from the release review (the regression tests use synthesized anchors re-encoded under the current crate, so they could not have caught a layout drift against the deployed bytes).

Side note captured in the second commit: fullnode.testnet.sui.io:443 currently 404s JSON-RPC, so the tools now carry an ordered endpoint list with a public fallback and skip-with-reason instead of panicking.

🤖 Generated with Claude Code