feat: emission scoring — capacity, volume, credibility ramp

[LandynDev]

Problem → Solution

Problem 1 — Idle crown farming. A miner can post just-above-market rates, self-reserve to lock themselves out, and earn full emission for "being available" without ever serving swaps. We can't simply stop paying reserved miners (inverse exploit: attackers force-reserve real miners to zero their emissions).

→ Capacity + volume weighting on per-direction rewards. Independent multipliers on the existing crown-time × success_rate³ formula:

reward_i = Σ_dir [pool_d × crown_share_id × sr_i³ × capacity_i] × volume_factor_i

• capacity_factor = min(1.0, collateral / max_swap_amount). A miner at the 0.1 TAO collateral floor with a 0.5 TAO max swap only covers 20% of the swap-size band, so they earn 20% of what they'd otherwise win. Cold-start fail-safe to 1.0 when bounds aren't readable.
• volume_factor = (1 - α) + α × min(1.0, vol_share / crown_share), α=0.5. Idle crown holders lose half; over-serving is capped at 1.0 — the cap is the anti-wash-trade guarantee (no incentive to force volume past parity with your crown share). Quiet network (zero volume) → factor 1.0, no penalty.

Each shortfall recycles to RECYCLE_UID; nothing transfers between miners.

Problem 2 — New-miner free emission. success_rate(stats) defaulted to 1.0 for zero-outcome miners, so a fresh hotkey holding crown earned the full cubed multiplier (1.0³ = 1.0) before serving anything. Cheap sybil farming surface.

→ Linear credibility ramp.

success_rate = (completed / total) × min(1.0, total / CREDIBILITY_RAMP_OBSERVATIONS)

CREDIBILITY_RAMP_OBSERVATIONS = 10. Zero observations → zero credibility. Each closed swap (completed or timed-out) climbs the ramp by 10%. At ≥10 closed swaps the ramp is a no-op and the function returns the raw success rate exactly as before. A miner with 8/10 still gets 0.8³ = 0.512 — no behavior change at steady state, only the bootstrap window changes.

This is the right shape for the network: it incentivizes brand-new miners to actually compete for crown and serve swaps (each successful fulfillment is a measurable credibility gain), instead of dropping in and farming from day one.

Recycling

All four shortfall sources flow to RECYCLE_UID:

• Success rate (sr³ < 1)
• Capacity (collateral / max_swap < 1)
• Volume (vol_share / crown_share < 1)
• Crown unfilled (no qualifying miner)

None redirect to other miners. Pool conservation is invariant (rewards.sum() == 1.0 always).

Surfacing

The per-round scoring trace now reads per crown earner:

uid=12 hotkey=5C1a.. crown_blk=400 sr=0.500 (5/10 closed, ramp=0.50) cap=0.20 (col=0.1t) vol=0.000t vol_share=0.00 crown_share=1.00 vol_f=0.50 reward=0.013

A miner sees exactly which factor zeroed them: collateral too low → raise; volume too low → fulfill more; closed swaps too few → keep grinding.

Test plan

• ruff check allways/ tests/ — clean
• ruff format --check — all formatted
• pytest tests/ — 461/461 passing locally (97 in scoring, including 19 new for capacity/volume/credibility behavior)
• Algebraic spot-checks on every new factor (helpers + integration)
• Pool-conservation regression: rewards.sum() == 1.0 across mixed-factor scenarios
• Schema migration is idempotent (duplicate column caught)
• Cold-start fail-safe verified: bounds_cache.max_swap_amount() returns 0 → capacity_factor = 1.0 (no zeroing on first pass post-restart)
• After merge: pull on lena, restart aw-vali, watch the scoring trace surface the new factors on the next pass