Wire owasp's score() to detect-only mode (fixes #30, #28 owasp half)

[alanzabihi]

Summary

• Implements real score() for the owasp benchmark adapter, replacing the NotImplementedBenchmarkError stub. Grades a contender's confirmed/rejected classification for one test case against expectedresults-1.2.csv ground truth into TP/FP/FN/TN (scoreOwaspVerdict in src/benchmarks/owasp/adapter.ts).
• standUpTarget() now requests autobrin-flue's detect-only mode (superagent-ai/autobrin-flue#182, merged into staging) via a new optional TargetHandle.detectOnly field, forwarded into the engagement payload by buildRepoPayload() (src/contenders/autobrin.ts). Additive/optional -- every other caller that doesn't set it is unaffected (existing buildRepoPayload tests unchanged).
• youdenIndex()/aggregateOracleScores() (src/oracle/types.ts) already report per-contender totals generically in src/matrix/report.ts's scorecard -- no OWASP-specific reporting code was needed once score() returns real OracleScores.
• PITHOS needed no adapter-side changes: it never takes a detectOnly-shaped payload, and its own pipeline already produces a confirmed/false_positive/inconclusive verdict independent of AutoBrin's stage machinery. Verified live (see below).
• Fixes cybergym/owasp score() error messages say capabilities are unmerged, but they merged today #28's OWASP half by removing the stale NotImplementedBenchmarkError message entirely (score() no longer throws), and removes the owasp row from BENCHMARK_CAPABILITY_DEPENDENCIES / AGENTS.md's capability table, matching how cve-bench was retired from that table once its dependency shipped.

Design note: location-aware matching (found via live verification)

scoreOwaspVerdict treats a confirmed finding as relevant to a task if its location overlaps the task's javaSourcePath/testName, or if it has no location at all. That fallback is load-bearing: AutoBrin's detect-only mode stops right after the adversarial gate, before the exploitation/disclosure stages that would otherwise populate a finding's location, so every AutoBrin ConfirmedFinding under this mode has location: undefined. PITHOS's findings, by contrast, do carry real file paths.

A live run surfaced exactly why the location check matters: scoring BenchmarkTest00063 (a path-traversal FP-trap, not vulnerable), PITHOS -- which scans the whole ~2,740-file vendored repo, not just that task's one servlet -- correctly found real but unrelated vulnerabilities elsewhere in the Benchmark's own test harness (hardcoded LDAP/keystore passwords). Without the location check, that would have been misattributed as a false positive for an unrelated task. See src/benchmarks/owasp/README.md's "Scoring" section and the corresponding test in tests/owasp.test.ts.

Real live verification

Ran 3 representative tasks (spanning pathtraver/CWE-22 and sqli/CWE-89, both vulnerable: true and false) end-to-end with both contenders via bench run owasp --contender <id> --model <model> --task <id>, dotenvx-injected real credentials, contributors: 1 to keep cost bounded:

Task	Ground truth	autobrin@staging (kimi-azure/kimi-k2.6, detectOnly)	pithos (kimi-k2.6, azure-openai-responses)
BenchmarkTest00001 (pathtraver)	vulnerable	rejected -> FN ($1.55, 811s)	rate-limited to an empty claim -> FN (non-crashing, but not a meaningful signal; see below)
BenchmarkTest00063 (pathtraver)	not vulnerable	rejected -> TN ($1.65, 778s)	confirmed 6 real-but-unrelated findings -> TN after the location fix above (555s)
BenchmarkTest00008 (sqli)	vulnerable	rejected -> FN ($0.87, 436s)	confirmed a real SQL-injection finding that names BenchmarkTest00008.java -> TP (1898s)

All 6 runs completed with real, non-crashing OracleScores -- no adapter/scoring code failures. AutoBrin's tracked cost across the 3 runs above: $4.07 (plus one earlier 3-contributor calibration run killed after $3.25 once --contributors 1 proved to cut cost ~3x with equivalent signal). PITHOS's cost is untracked by the harness (costStatus: 'unavailable', pre-existing for this contender), so no dollar figure for those runs.

Two infrastructure findings surfaced by this verification, neither a bug in this PR's code, reported here for visibility:

• PITHOS's --ref can't resolve a raw commit SHA. contenders/pithos.ts's buildPithosArgs passes target.sha straight through as --ref, which PITHOS's own CLI turns into git clone --branch <ref>. That works for branch/tag refs (e.g. repo-cve-smoke's pinned tags) but fails for OWASP's pinned raw commit SHA ("Remote branch <sha> not found in upstream origin"), despite PITHOS's own --help text claiming commit support. Worked around locally for verification by pre-seeding PITHOS's repo cache from our own vendor clone at the pinned commit; not fixed here since it's PITHOS's own CLI, not this repo.
• Shared Azure Kimi-K2.6/eastus2 capacity is contended. Multiple parallel subagents were running concurrent kimi-azure engagements against the same deployment while this PR's live verification ran. AutoBrin's own provider-503/retry handling absorbed this fine; PITHOS's retry backoff (3 attempts, 2s/4s/8s) gave up faster under sustained 429s and fell back to empty "repair" results for BenchmarkTest00001 across 3 attempts. Not a bug in this PR; flagging in case it affects other concurrent OWASP/PITHOS verification.

Test plan

• npm run validate (typecheck + 209 tests, all green)
• Local Bugbot review of the uncommitted diff: no bugs found
• Real live verification above (3 tasks x 2 contenders, all non-crashing)

Fixes #30. Fixes the OWASP half of #28 (CyberGym half handled by a separate PR).

[cursor]

Current version of PR was reviewed by /review-bugbot on Jul 1, 23:35 GMT+2. It flagged 0 findings.

^{Bugbot on commit 59e7c39 is skipped.}