fix(pm): round-robin batch-epoch settle scan + idle fast-path (follow-up to #139)

[chiliec]

Summary

Closes the two gaps left open in the #139 review (stacked on #139 — its commit is included; merge #139 first or merge this and it lands too):

1. Idle fast-path. After fix(pm): stop batch-epoch settle from starving markets behind the processing cap #139 the boundary scan no longer let idle markets eat the cap, but each idle LMSR market still paid the q-vector snapshot (outcome_count index finds) before its idleness was detected. The bet-index probe now runs first: an idle market costs exactly one lower_bound. This is safe because bets at a market's current epoch are queued-only — settled/refunded bets always live at past epochs — so an empty probe reliably means idle.

2. Round-robin cursor. The scan always restarted at the lowest market id, so ≥ cap always-busy low-id markets would still deterministically starve newer ones — and that state is cheap to manufacture (dust batch bets on the oldest cap markets every epoch). The scan now resumes from pm_batch_settle_cursor (new consensus field on dynamic_global_property_object, same pattern as the pm seed flags), wrapping at most once per boundary. Cursor = next unvisited active market when the cap cuts the scan short, 0 after a full circle.

The remaining ceiling is noted in-code: the walk is O(active batch markets) per boundary at one probe each; if that ever hurts, index queued bets by (status, market) and drive the scan from that.

Index change

pm_market_index by_status goes from ordered_non_unique(status) to ordered_unique(status, id) so the cursor can seek mid-range. All existing users do prefix lookups (scalar lower_bound in the crons, equal_range(status) in prediction_market_api) and keep working; in-status iteration order becomes deterministic-by-id instead of insertion order. Requires a chain reset — as does #139 itself (its behavior change is ungated), so the pm testnet reset covers both.

Tests

Two consensus_sim regression scenarios (BUILD_CONSENSUS_TESTS), both designed to fail on the code they guard:

• batch_settle_idle_markets_do_not_starve — cap=1, three allow_batch markets, a queued bet only on the newest; pre-fix(pm): stop batch-epoch settle from starving markets behind the processing cap #139 code starves it forever and bumps idle epochs (both asserted).
• batch_settle_round_robin_prevents_busy_starvation — cap=1, market 0 re-fed a queued bet before every epoch boundary; without the cursor market 1 never settles. Block phasing is arranged so market 0 is queued at every boundary (otherwise the idle-skip alone would mask the cursor).

Verification

• Syntax-checked pm_evaluator.cpp, prediction_market_api.cpp, and test_pm_lifecycle.cpp against the pm-branch generated hardfork header (clang, -fsyntax-only, clean).
• Not linked or executed locally (arm64 host cannot link equihash); the consensus_sim scenarios need an x86-64 build with BUILD_CONSENSUS_TESTS=ON — same CI gap already tracked from the HF14: Prediction Markets #124 review.

🤖 Generated with Claude Code