feat(bounty): ServiceBounty + evaluator market — demand-side bounties with commit-reveal evaluation

[bussyjd]

Stack position

PR 1 of 2 in the ServiceBounty stack. The real-money escrow leg follows in a stacked PR based on this branch. Independent of #632/#633 (skill marketplace / smoke agent) except for shared additive internal/erc8004 calldata builders — whichever merges second rebases trivially.

What

A demand-side marketplace primitive: a poster publishes a ServiceBounty CR ("do X for reward R"), a fulfiller agent claims and submits, and an evaluator market decides whether the work passes — with payment held in an escrow seam until the verdict.

Core pieces:

• ServiceBounty CRD + controller (internal/serviceoffercontroller/bounty*.go): lifecycle Open → Claimed → Submitted → Evaluating → Paid/Rejected, driven by annotation write-channels so agents interact with plain kubectl RBAC, never controller credentials.
• Verification by default: spec.eval.mode gates payout on an evaluator verdict; opting out requires the explicit --dangerously-skip-verification flag.
• Commit-reveal evaluation: a panel of k evaluators commits sha256(score|salt|address) hashes, then reveals; median-of-k is the verdict; non-reveals are penalized as outliers. Quorum, reveal windows, and outlier bands are spec'd per bounty with sane defaults.
• Evaluator ladder (EvaluatorEnrollment CRD): Shadow → Probation → Full progression — shadow evaluators score without weight, probation carries a value cap and half pay, divergence from the median sets careers back. No staking, no slashing; sybil cost comes from the unpaid shadow period and self-bonds.
• Self-bonds: fulfillers attach a bond forfeited on rejected work (escrowed alongside the reward).
• A2UI reports: bounty results render as structured agent-to-UI report documents (v1.0-candidate schema) incl. a bounty_report MCP tool over the existing paid-MCP seam.
• Task-type registry: dynamic, versioned task packages (benchlocal@v1 enabled; finetune@v1 staged behind enabled:false).
• ERC-8004 wiring (internal/erc8004/{reputation,validation}.go): calldata builders + readers for the Reputation/Validation registries — evaluator verdicts can be grounded on-chain (operator-submitted; the controller never signs).
• Poster/fulfiller/evaluator CLI: obol bounty create|claim|submit|eval commit|eval reveal|status|... plus calldata derivation commands.

Security invariants (test-pinned)

• Bounty reconcile creates no HTTPRoute/Middleware/ReferenceGrant/Secret/Namespace — the controller's blast radius doesn't grow.
• Controller is read-only on evaluator enrollment specs.
• Agent RBAC additions are namespace-scoped; admission hardening pins what agent SAs may write.
• CRD ↔ Go parity test (caught a real pruning bug during development — kept as regression).
• Escrow URL/credentials reach the controller via env only, never via CR spec/annotations.

Why a squash commit

This squashes a 15-commit development series (list in the commit message) so the PR reviews as one coherent unit. Granular history is preserved locally and can be pushed on request.

Validation

Full unit + controller-test suite green (panel selection determinism, commit-reveal verdicts, ladder transitions, parity, admission). Design docs included under plans/ (bounty-ane-marketplace-design.md, evaluator-market-research-notes.md). The escrow PR stacked on top carries the end-to-end money-leg validation.

🤖 Generated with Claude Code

[OisinKyne]

We don't have any buyers with unfilled demands. What types of services do you anticipate here?

I don't think any of our buyers and sellers need an escrow and evaluator yet. I suggest leaving this unmerged until I have a better understanding of who needs this