feat(scorer): add reason sub-label to profit_scored_total

[Pablosinyores]

Summary

`aether_mempool_profit_scored_total` was labelled only by `decision`, so
the dashboard could see "10 reverted" but not whether they came from
the V2 U256 walker (PR #136), the f64 absurdity floor (PR #136), the
V3 revm verifier (PR #144), or organic revm reverts. Add a `reason`
sub-label distinguishing those code paths so PR #146's decision-pie

• scored-rate panels can split-by-reason.

Stacks on PR #148 (`feat/dedupe-replay-scorer-helpers`).

Five reason values

Wire labels pinned by `reason_constants_are_stable_wire_labels`:

Reason	When it fires	Pairs with
`n/a`	non-reverted decisions, `no_path`, or any path with no sub-source worth distinguishing	any
`u256_walker`	V2-only exact-U256 walker (`verify_cycle_u256`) reached a verdict (PR #136)	profitable / unprofitable / reverted
`absurdity_floor`	f64 fallback above `MAX_PLAUSIBLE_F64_NET_WEI` (1 ETH) downgraded to reverted (PR #136)	reverted
`revm_verdict`	V3-touching revm sim (`verify_cycle_revm`) ran to completion with a non-reverting verdict (PR #144)	profitable / unprofitable
`revm_revert`	V3-touching revm sim explicitly reverted/halted (PR #144)	reverted

Wire-format

• Metric: `aether_mempool_profit_scored_total{decision, reason}` —
  cardinality bounded at ~10 `(decision, reason)` combinations.
• Persistence: `reason` is Prometheus-only. It does NOT flow into
  the `mempool_profitability` Postgres table because the migration's
  CHECK constraint only covers `decision`. `NewProfitabilityScore.reason`
  is consumed by `PgProfitabilityWriter::insert_score` for the metric
  label and dropped before the SQL bind. No migration needed.

Code changes

`crates/grpc-server/src/profitability_writer.rs` (+62 / −5):

• New `REASON_*` `pub const` block.
• `scored_total` `IntCounterVec` bumped from `["decision"]` to
  `["decision", "reason"]`.
• `NewProfitabilityScore` gains `pub reason: &'static str` with a
  `serde(default = "default_reason")` shim returning `REASON_NA` so
  external callers that haven't been updated still deserialise cleanly.
• `PgProfitabilityWriter::insert_score` reads `score.reason` before
  the `try_send` and uses it in the metric `with_label_values` call.

`crates/grpc-server/src/bin/aether_profit_scorer.rs` (+45 / −19):

• `ScoreOutcome` gains `reason: &'static str`.
• `revm_verdict_to_decision` and `f64_fallback_verdict` return a
  4-tuple `(net, realised, decision, reason)`.
• The aggregating `let` in `score_one` destructures the 4-tuple and
  the U256-walker arm explicitly attaches `REASON_U256_WALKER`.
• `no_path_outcome` carries `REASON_NA`.
• `score_batch` plumbs `score.reason` into the
  `NewProfitabilityScore` constructed for the writer.

`deploy/docker/grafana/dashboards/mempool.json` (+8 / −8):

• Panel fix(executor): implement real eth_sendBundle builder submission #19 (Decision breakdown): `sum by (decision, reason)`,
  legend `{{decision}} / {{reason}}`, title and description updated.
• Panel fix(executor): wire gas oracle and nonce manager to real RPC #20 (Scored rate): same change.
• Other panels untouched.

Verification

```
cargo clippy --workspace --all-targets -- -D warnings ✓ clean
cargo test --workspace --lib --bins ✓ 528 passed, 0 failed
python3 -m json.tool mempool.json ✓ parses
```

New unit test:

• `reason_constants_are_stable_wire_labels` (pinned wire strings)

Existing verdict-helper tests gain `reason` assertions:

• `revm_verdict_decision_mapping_reverted` → asserts `REASON_REVM_REVERT`
• `revm_verdict_decision_mapping_profitable` → `REASON_REVM_VERDICT`
• `revm_verdict_decision_mapping_unprofitable` → `REASON_REVM_VERDICT`
• `f64_fallback_verdict_above_floor_reverted` → `REASON_ABSURDITY_FLOOR`
• `f64_fallback_verdict_below_floor_profitable` → `REASON_NA`
• `f64_fallback_verdict_negative_unprofitable` → `REASON_NA`
• `no_path_outcome_carries_negative_net_when_gas_given` → `REASON_NA`
• `metrics_register_round_trips` emits across multiple (decision, reason) pairs.

Operational notes

After this lands, the live scorer process (PID 2476 at handoff time)
needs to restart to pick up the new metric label set. Grafana caches
metric metadata; the panels will start splitting once the new metric
samples arrive. The pre-restart counter series (cardinality
`{decision}` only) coexists with the new `{decision, reason}` series
without breakage — Prometheus treats them as distinct families.

Out of scope

• `no_path` sub-reasons. The handful of upstream causes
  (pool absent, eth_call empty, optimiser bailed, tokens missing,
  zero reserves) all collapse to `REASON_NA` today. If one proves
  operationally interesting (e.g. sustained `pool_not_in_registry`
  signals a V3 pool-discovery gap), it can be split out in a follow-up.
• Counter cardinality bound. With 4 decisions × 5 reasons = 20 max
  combinations, the cardinality is well below any Prometheus practical
  limit. No `relabel_config` needed.

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
aether	Ready	Preview, Comment	May 20, 2026 12:20pm
aether-63xv	Ready	Preview, Comment	May 20, 2026 12:20pm