jacobian: v0.10.0

0.10.0 (2026-08-08)

Features

• benchmarks: add fail-closed aggregate reward policy and ratchets (70963ed)
• benchmarks: add selected-task workflow (#572) (9d94310)
• benchmarks: move verifier exceptions into task-local contracts (ccf4b7c)

Bug Fixes

• address post-merge MCP and observation review findings (7f5a035)
• address post-merge review findings (9913af4)
• benchmarks: accept compact network_mode TOML assignments (#668) (5957cd2)
• benchmarks: address reconstruction deck certificate review comments (#657) (7154612)
• benchmarks: bind cddlib/regina and tighten provider Oracle evidence (#709) (bf44e7a)
• benchmarks: discard consumed streaming whitespace prefix in reconstruction deck verifier (#682) (256ff13)
• benchmarks: discover conjecture host validations (#829) (08f7d95)
• benchmarks: fail-closed reward, support, and provider evidence (f03baaf)
• benchmarks: identify provider feasibility tasks accurately (#771) (4de8452)
• benchmarks: install jsonschema in upgraded verifier images (64132a9)
• benchmarks: land Hadwiger verifier review follow-ups (evidence schema, JSON type fidelity, assurance guard, non-finite JSON) (#675) (c174872)
• benchmarks: make mathematical rewards fail closed (b3dbc1c)
• benchmarks: migrate legacy verifier diagnostics (ee14bf7)
• benchmarks: pin deterministic pytest-randomly seed for host shards (#625) (6c893d3)
• benchmarks: preserve decoupled input diagnostics (9b05caf)
• benchmarks: preserve fail-closed diagnostics and validation gates (e2c264b)
• benchmarks: publish bound input for provider Oracle verifiers (#683) (41504be)
• benchmarks: repair shared CI root causes for oracle and contracts (#640) (95b2024)
• benchmarks: require authoritative completion counts (556dfae)
• benchmarks: skip vanished Oracle evidence candidates (#743) (a3b35ac)
• benchmarks: sort functools import for schema validator cache CI (#621) (db4cc38)
• capabilities: return typed failures for invalid adapter output (#739) (166a380)
• checkers: bind the Lake launcher digest to the authorized Lean runtime (b84a9ca)
• ci: account for shared Lean process memory (6114242)
• ci: avoid repeated trajectory cluster validation (a7c312b)
• ci: format eval telemetry test signature (#777) (bad5a78)
• ci: restore lint-full and test-architecture on main (0891065)
• ci: restore lint-full and test-architecture on main (f39f3eb)
• ci: restore static gates after benchmark additions (#723) (b5fc67a)
• ci: tolerate duplicate timing entries across benchmark shards (#624) (6b3b58a)
• cli: preserve command failures during cleanup (#713) (fb118bb)
• eval: guard resource-read tool field against non-hashable values (2e08f67)
• eval: ignore malformed telemetry status and tool fields (#708) (f64d891)
• experiments: use first-page parent capacity (#716) (10421cb)
• fail closed on malformed JSON and purge stale plugin imports (4cdb2b4)
• fail closed on malformed observation/control JSON and purge stale plugin imports (3bfaf5f)
• fail-closed evaluation integrity across verifiers and kernel (#585) (5aec651)
• lean: resolve toolchain lean instead of elan proxy (#627) (c91140e)
• matrix: reject invalid flint worker protocols (#725) (f2d3525)
• MCP concurrency, fail-closed benchmark, and resource safety fixes (ed07ff7)
• MCP concurrency, fail-closed benchmark, and resource safety fixes (6dab752)
• mcp: bound cancellation drain with monotonic deadline (fec2f27)
• mcp: clarify validation and assurance diagnostics (#602) (0733334)
• mcp: distinguish unknown discovery domains (#604) (fcfe72f)
• mcp: quarantine failed tenant runtimes and serialize shutdown (#745) (ce452d9)
• mcp: report nested required constraints, keep capability selection agent-owned (#678) (25141d0)
• npm: resolve latest bootstrap for upgrades (fb884b3)
• polynomials: bound inverse solver termination (#744) (024d9c7)
• pre-commit: remove stale verifier-support hook (#730) (383d17e)
• providers: correct Z3 mismatch diagnostic (#748) (0f7a445)
• runtime: attempt all owners during shutdown (#714) (b423a41)
• runtime: preserve bootstrap failures during cleanup (#724) (42fc483)
• search: bound nomination lineage before persistence (#750) (f6043fe)
• smt: allow declare-const in cvc5 worker (#715) (92ca13f)
• smt: reject non-string cvc5 worker statuses (#726) (43ceea8)
• tests: audit and clean up false greens, assertion debt, and antipatterns (#628) (8f79a00)
• tests: audit and clean up test anti-patterns (#641) (7eafbea)
• tests: avoid sharing template blob inodes (2d3d260)
• tests: avoid sharing template blob inodes (a729e16)
• verification: gate VERIFIED on TRUE and fail closed on unknown ops (721635e)
• verification: include supporting artifact metadata (#711) (6714453)
• verification: verify accepted false conclusions (d478e2e)
• verifier: retain false-certification diagnostics (#773) (a3cbb91)
• verify accepted false conclusions (dfcc29c)

Performance Improvements

• benchmarks: cache schema validators with lru_cache (#626) (f55147f)
• schema: skip meta-schema validation for Pydantic-generated schemas (c928039)
• schema: skip meta-schema validation for Pydantic-generated schemas (18d6042)
• storage: reuse prepared artifact identities (#747) (d43167a)
• tests: hardlink immutable blobs in template copy (65b0092)
• tests: hardlink immutable blobs in template copy instead of copying (26fc5af)

Documentation

• add comprehensive audit report (e42392b)
• agents: define mathematical interoperability rules (#769) (3097458)
• benchmarks: document prebuilt agent environment direction (e24dbf6)
• clarify inline verification evidence (53bd6aa)
• clarify mathematical agent workflow (033bb2b)
• explain why atomic capabilities scale (a9df893)
• skill: capture benchmark validation lessons (294e38f)
• skills: capture verifier validation lessons (ec56113)