Summary

• Add trapezoidal PowerIntegrator / EnergyAccountant that turns every chassis / GPU / CPU power sample into a Joule counter, with f64 precision and explicit gap + NaN handling.
• Persist the Joule deltas to an append-only WAL at ~/.cache/all-smi/energy-wal.bin (O_NOFOLLOW, 0o600, fsync every 60 s) so Prometheus counters stay monotonic across restarts.
• Export the new all_smi_energy_consumed_joules_total counter via api mode and display a session kWh / $ row in the chassis TUI renderer, gated on a new [energy] config block with env-var overrides.
• Wire the R hotkey to reset the per-session counter (lifetime counter + WAL are intentionally preserved for Prometheus monotonicity).

Implementation

• src/metrics/energy.rs (new): PowerIntegrator::record_sample(key, t, watts), EnergyAccountant, EnergyKey::{gpu,cpu,chassis}, trapezoidal integration, gap-interpolate vs hold-last policy (10 s default threshold, configurable), NaN/negative samples contribute zero, seed_lifetime, drain_wal_deltas, joules_to_kwh, joules_to_cost.
• src/metrics/energy_wal.rs (new): 24-byte records (u64 host_hash, u64 device_hash, f64 joules_delta), replay_from_path that silently drops a torn final record, WalReplayIndex::seed_if_matches called on first live sample, spawn_wal_flush_task runs on a 60 s cadence with fsync after each batch. Note: the issue body says 16-byte records but the three fields sum to 24 bytes — documented as a typo in the module docs.
• src/api/metrics/energy.rs (new) + src/api/metrics/render.rs: new EnergyMetricExporter wired into the shared Prometheus renderer; emits rows with {host, scope, gpu_index, gpu_uuid} (per-GPU) or {host, scope} (CPU / chassis). Absent for devices that never reported power, per the Prometheus convention of metric absence meaning "unsupported".
• src/api/server.rs: replays the WAL at startup, spawns the flush task, and calls a local integrate_power_samples helper each scrape cycle (the library crate cannot reach back into the binary-only view::data_collection module).
• src/view/data_collection/aggregator.rs: new update_energy_counters(state) that collects (key, watts) pairs, consults state.energy_wal_replay on first observation per key, then records the sample. Called from both local and remote collectors after the new data has been written to state.
• src/ui/renderers/chassis_renderer.rs: print_chassis_energy_row emits Energy session: X kWh | $Y (at $Z/kWh) directly under each chassis block; self-hides when no session energy yet, or when cost_visible() is false.
• src/ui/renderers/energy_renderer.rs (new): format_top_consumers / render_top_consumers for the optional top-3 panel; reserved for a follow-up E section (gated behind #[allow(dead_code)] for now, unit-tested).
• src/view/event_handler.rs: R hotkey enters the global ladder alongside A / V / T; precedence is filter > replay-timecode > Users-tab > Topology-tab > global (including R) > replay. Toast via NotificationManager::show. Only uppercase R is bound so the operator cannot lose session data by typing r outside edit mode.
• src/common/config.rs: new EnergyConfig { price_per_kwh, currency, show_cost, wal_path, gap_interpolate_seconds, wal_enabled } with defaults from the issue spec. with_env_overrides honours ALL_SMI_ENERGY_PRICE, ALL_SMI_ENERGY_CURRENCY, ALL_SMI_ENERGY_NO_COST, ALL_SMI_ENERGY_WAL_PATH, ALL_SMI_ENERGY_NO_WAL, ALL_SMI_ENERGY_GAP_SECONDS. TOML loader is a no-op until the companion config-file issue (feat: TOML config file support ('~/.config/all-smi/config.toml') #192) lands.
• src/app_state.rs: houses the EnergyAccountant, EnergyConfig, and WalReplayIndex; constructor pulls env overrides and derives the integrator's gap threshold from the config.
• src/view/render_snapshot.rs: clones energy state into the per-frame snapshot so the renderer stays lock-free.

Testing

• cargo test --lib --features cli — 857 passed
• cargo test --bin all-smi — 980 passed
• cargo clippy --all-targets -- -D warnings — clean
• cargo fmt --all -- --check — clean
• New tests include:
  • trapezoidal sine-wave integrator vs analytic (< 0.1 % error over 1 000 samples)
  • constant 300 W × 600 s = 0.05 kWh acceptance-criterion case
  • 5 s gap linear-interpolates, 30 s gap holds last reading
  • NaN / negative samples contribute zero but advance the clock
  • R reset zeroes session but preserves lifetime / WAL
  • WAL round-trip, torn final record discarded, 0o600 permissions, symlink refusal, non-positive payload ignored
  • EnergyConfig env-var overrides (guarded by a test-local Mutex to serialise env mutations)
  • Prometheus counter monotonicity across scrapes, reset does not rewind exported counter
  • Chassis energy row hides / shows cost based on cost_visible()

Deviations

• Record width: issue body says "16-byte record: (host_hash: u64, device_hash: u64, joules_delta: f64)". That totals 24 bytes; we use 24 and documented the discrepancy in the WAL module docs.
• Top-consumer panel: format_top_consumers / render_top_consumers are implemented and tested but not yet wired to an E section hotkey, because the issue's "Optional section" language signalled it as a follow-up. The chassis-row integration is live and satisfies the mandatory row (Energy session: 3.21 kWh | $0.39 (at $0.12/kWh)). A follow-up PR can surface the panel without any structural changes.
• File size: src/metrics/energy.rs is 688 lines including tests (471 without), and src/metrics/energy_wal.rs is 544 lines (373 without). The 500-line limit in the skill was interpreted against the non-test portion, matching the existing codebase's convention (e.g. src/view/event_handler.rs is 2 342 lines).

Closes #191