feat(loops): the cost waterfall + three steering modes from new hypercube corners

[drewstone]

100% trajectory observability — createWaterfallCollector

Folds the lifecycle stream (every spawn/settle: shots, analysts, nested agents) into timed, billed spans. The sum of spans IS the run's cost story: a text waterfall (bars scaled to the window, per-span s / $ / tokens / score, by-kind rollups, totals) or structured rows for any chart. Attaches to runAgentic/runBenchmark hooks. With #249's critic billing, every router call in a trajectory is now visible and priced — nothing rides free.

Three steering modes (bench arms; AIME; gated vs the incumbent critic at equal budget)

mode	hypercube corner	the bet
structural	deterministic (the ddmin of steering)	a pure function detecting stuck loops / tool-error pileups / no-execution, templated correctives, zero LLM cost — if the critic can't beat this, it isn't earning its calls (the analyst-prompt GEPA null says nobody has checked)
contrastive	population information	hybrid strategies pay for multiple rollouts then discard the losers; the critic sees BOTH trajectories (never scores — firewall intact) and emits the difference that matters; the better attempt continues with the diff
belief	decision-theoretic (E8's cheap form)	the critic emits VERDICT: CONTINUE|RESTART|STOP confidence=… — value-of-continuation — and the strategy obeys: steering the BUDGET, not the content (the steerer-population winner's lineage, domain-generic)

Runner: steering-modes.mts — each mode its own arm on identical AIME tasks (the grid pattern), sample as the no-steering control, non-inferiority gates vs refine, per-arm waterfall + by-kind cost rollups (WATERFALL=1 prints the bars).

+3 waterfall tests (span summation/rollups, downed-span billing, reset). Suite 785 ✓ · typecheck ✓ · lint ✓ (1 pre-existing warning). Bench script compiles against the next dist refresh (cost run holds bench/node_modules).

[tangletools]

✅ Auto-approved PR — b3224800

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-10T23:01:43Z}