Add OpenCode as a third execution harness

[chenmingtang830]

Summary

• add OpenCode 1.18.3+ as a third open-source core harness for API, CLI, and SDK cells
• require an explicit provider/model route and scope only declared provider and target environment values
• decode OpenCode JSONL through the stable event seam from Refactor transcript decoding behind stable events #217, including normalized tool counts and token/cache metrics
• fail OpenCode MCP cells closed as unsupported before credentials, health checks, provisioning, or invocation
• preserve the historical ax.normalized-result/v1 enum while one-cell records carry unsupported-surface

Isolation and evidence

• run with --pure and --auto from a disposable cwd outside the checkout, with private HOME/XDG roots, no ambient auth, no MCP, and no task subagents
• block system/MDM-managed OpenCode config and reject pack-declared OpenCode/XDG control variables before binary detection
• delete the binary session database after recovery
• omit tool outputs, errors, and reasoning text from durable JSONL; exact-redact every harness-visible credential and pack value
• fail closed on malformed result shapes, missing trace evidence, unsafe short values, timeouts, and harness mismatch
• record a bounded controller-derived outcome reason for every retry attempt
• preserve --task in the generated verify command so single-task pilots use the correct denominator
• reject objectively observed vendor-CLI actions in an API cell as cross-surface evidence

Live pilots

• OpenCode 1.18.3 with opencode/deepseek-v4-flash-free
• direct smoke emitted native JSONL and wrote the exact requested file; it caught that 1.18.3 rejects --no-env-file, fixed in e7d5ac3
• Exa example-pack API task made a real Search call and passed live Contents read-back: 1/1, 100%; updated normalization reports 7 tool calls across 28 events
• InsForge DAEB research task created and inspected the exact run-scoped table; arena SQL read-back found both expected fields and bounded reset deleted 1/1 resource
• the DAEB transcript also used npx @insforge/cli in an API cell; updated verification now reports this as cross-surface FAIL instead of a clean API pass, and reports 34 tool calls across 103 events
• DAEB remains research evidence only: this generic integration does not add OpenCode to the canonical arena production policy or publication matrix

Validation

• core Vitest: 58 files, 540 tests passed
• arena Vitest: 47 files, 285 tests passed with one worker (default macOS parallel run hit process-launch timeouts; Linux CI is authoritative)
• npm run typecheck
• npm run build
• npm --cache .npm-cache pack --dry-run
• real Exa single-task live verification: 100% (1/1), gate passed
• DAEB transcript replay: correct 1-task denominator, tool count, and cross-surface failure

Stacked on #217; merge that PR first.

[github-actions]

Normalized records diff

✅ PASS — No correctness regression detected

Compared 4 baseline surface cell(s); no existing pass@1 or pass³ score decreased or disappeared.

Overall is a macro-average of participating surface scores. Agents are never averaged together.

Overall by agent

Product	Agent	Base pass@1	Head pass@1	Δ	Base pass³	Head pass³	Surfaces
fixture-db	claude-code	92.9%	92.9%	+0.0 pp	78.6% (11/14)	78.6% (11/14)	2
fixture-db	codex	78.6%	78.6%	+0.0 pp	64.3% (9/14)	64.3% (9/14)	2

Surface cells

Product	Agent	Surface	Base pass@1	Head pass@1	Δ	Base pass³	Head pass³	Latency ms	Cost USD	Tokens	Harness version
fixture-db	claude-code	api	85.7%	85.7%	+0.0 pp	71.4% (5/7)	71.4% (5/7)	1100 → 1100	0.4200 → 0.4200	3350 → 3350	2.1.5 → 2.1.5
fixture-db	claude-code	cli	100.0%	100.0%	+0.0 pp	85.7% (6/7)	85.7% (6/7)	1300 → 1300	0.5100 → 0.5100	3720 → 3720	2.1.5 → 2.1.5
fixture-db	codex	api	85.7%	85.7%	+0.0 pp	71.4% (5/7)	71.4% (5/7)	1200 → 1200	— → —	3600 → 3600	0.121.0 → 0.121.0
fixture-db	codex	cli	71.4%	71.4%	+0.0 pp	57.1% (4/7)	57.1% (4/7)	1500 → 1500	— → —	4100 → 4100	0.121.0 → 0.121.0

Operational metrics are context only and never affect the verdict or ranking.