docs: publish Ollama DeepSeek Flash harness comparison

[Astro-Han]

Summary

• Publishes the audited 89-task Maka versus OpenCode run on Ollama Cloud deepseek-v4-flash:0731, including the per-task outcome projection.
• Combines this result with the previous DeepSeek V4 Flash run using a task-clustered exact sign-flip test: Maka leads 113 to 92 across the two runs, with p = 0.00416.
• Separates effectiveness from economics: the current run is 52 to 43, while OpenCode uses fewer tokens on the 88 fully metered pairs.
• Discloses the one non-recoverable final-usage gap, budget-exhaustion shift, provider-condition difference, and full infrastructure audit. No harness code changes are included.

Follow-up to #1719.

Verification

• git diff --cached --check
• Recomputed the committed CSV aggregates: 89 tasks, Maka 52, OpenCode 43, 18 Maka-only, 9 OpenCode-only, 25 versus 18 budget exhaustions.
• Recomputed the two-run task-clustered exact result from both committed CSVs: 21-outcome lead, p = 0.004163928097113967.
• Verified the committed CSV SHA-256 and all relative Markdown links.
• Audited all 178 final Harbor results, 176 CTRF reports, both missing-CTRF traces, verifier retry evidence, and OpenCode provider telemetry; no accepted infrastructure failure was found.
• Code tests were not run because this PR changes only Markdown and CSV evidence.