fix(benchmarks): remove hidden Farkas evidence cap

[kaoru0822-kitauji]

Summary

This is one bounded benchmark-integrity change based on an audit of historical PR #876 against the current Harbor architecture at 413e3846c2529af9e8f90a96d6f5679b23ed04f2.

The audited Farkas task accepted schema-valid evidence with arbitrary JSON whitespace in its public contract, but the hidden verifier rejected files above 1 MiB. This removes that unpublished cap and ports the task-local loader to the repository's existing streaming JSON pattern. No product/runtime behavior, evaluator strategy, or general agent skill is added.

Integrity audit

Invariant	Current assessment
Hidden task material is isolated	Already solved: Harbor runs agents and verifiers in separate Docker environments, with static guards against hidden copies, widened build contexts, visibility leaks, and symlink escapes.
Revisions are reachable and immutable	Already solved: source SHA, clean snapshot identity, task digests, and observation records bind current reachable Git trees.
Terminal labels do not depend on a reasoning-log protocol	Historical/obsolete: terminal states come from the exact task verifier; telemetry is passive and has no verdict authority.
Evidence has no hidden size limit	Reproducibly broken here: exact-farkas-ldl-slice had a hidden 1 MiB cap. This PR fixes that one task.
Equivalent encodings are accepted	Solved for the audited tasks through explicit canonical rules or semantic comparison.
Limits are task-owned	Solved: Harbor task limits live in task.toml; timeout/error states remain non-conclusions.
Every validity rule is public	Partially solved before this PR because the Farkas size cap was unpublished; fixed for this task.
Alternate valid witnesses are accepted	Solved in the sampled tasks (including alternate LDL/Sylvester evidence, RP2 witnesses, primes, and prose).
Malformed/false assurance fails closed	Solved through task-local and generic adversarial matrices.
Local versus generic verifier ownership is clear	Solved: task semantics stay local; generic metadata/integrity attacks remain in the shared matrix.

Frozen task sample

Source SHA: 413e3846c2529af9e8f90a96d6f5679b23ed04f2

• exact-farkas-ldl-slice: sha256:5ff1eed2a5ff443222d181f3c676ac5012b4cd01189f1d983ef2e5256e447848
• polynomial-map-collision: sha256:9243ca8ee4800cb8eae039a08e0414c7fcfd8e111c3d1a3c8d6fbbfc44884abd
• rp2-homology-lattice: sha256:35c60550453aa1979b4694a399813023844ab3a6524643e106a9e983836b6539
• apollonius-gap-repair: sha256:701b5fc809ec6973a55d8a30568d89829c9c7f683b80c04241a4df2d6bc61636
• elementwise-fixed-no-global-invariant: sha256:ce3612dd2ebff71eb14234c965ef1d8959e19e43f0953b3366df13f04446f2ae

Prospective changed-task digest: sha256:174ddb8eb0f35a81f21a196aaa7dc562f5af399278da763b5bc6f7f5c0bf2e58.

Pilot evidence

The planned paired pilot was gpt-5.6-luna at medium reasoning, with a no-Jacobian control and a current-Jacobian-MCP treatment, and no general evaluation skill exposed. Both substantive attempts ran in named tmux sessions.

The host has no Docker daemon, so a Bubblewrap clean room was used after proving that the repository and verifier files were not visible. The attempt was stopped after repeated infrastructure failures: the installed standalone Codex package lacked its code-mode host executable inside the clean room, so the model could not read or write the task workspace. The control produced only a blocker trajectory (59,307 input, 52,992 cached input, 552 output, 142 reasoning tokens); neither condition produced a mathematical submission. The comparison is therefore INCONCLUSIVE, not a treatment result. It provides no evidence for adding a general tool-use skill, and this PR adds none.

Change

• Remove the unpublished MAX_EVIDENCE_BYTES rejection.
• Stream one JSON value while discarding arbitrarily large legal leading/trailing whitespace.
• Continue to reject leading/trailing non-whitespace garbage and duplicate JSON values.
• Add a regression above 1 MiB plus adversarial garbage cases.
• Refresh the task verifier checksum.

Validation

• make harbor-plan BASE=origin/main: selects only the Farkas host leaf, its generic verifier-contract leaf, and its Oracle.
• Focused task leaf: 10 passed.
• Generic verifier-contract leaf: 12 passed, 1,658 deselected.
• Static task quality and contract checks: passed.
• git diff --check origin/main...HEAD: passed.
• make check: Ruff, formatting, complexity, mypy, and 869/870 unit tests passed. The remaining unrelated test_python_distribution_identity_binds_installed_file_bytes failed due same-timestamp metadata caching; its isolated reproduction failed once and passed immediately on retry without a code change, and the full retry hit the same pre-existing timing condition.
• Exact Harbor Oracle: not run because Docker is absent (/bin/sh: docker: not found), so no Oracle verdict is claimed.

Deferred proof gaps

A Docker-capable host should run the selected exact-task Oracle and the frozen five-task paired pilot. Those are explicit evidence gaps; they do not broaden this PR beyond the single reproducible verifier-contract defect.