fix(providers): inject top-level instructions for openai-responses wire (#134)

[hqhq1025]

Summary

• Strict OpenAI-Responses gateways (sub2api-style routers) return 400 when input[] carries a system/developer role without a matching top-level instructions field. pi-ai's plain openai-responses wire emits the former but not the latter.
• Mirror the openai-codex-responses wire's strict behavior via pi-ai's onPayload hook: set params.instructions from the aggregated systemPrompt, and filter out role === "system" | "developer" entries from input[].
• Only wired when model.api === "openai-responses" AND systemPrompt is non-empty. Other wires (anthropic-messages, openai-completions, openai-codex-responses) are untouched.

Fixes #134.

Four principles

• Compatibility: green — no schema or IPC change; only the wire payload for openai-responses is adjusted; other wires unchanged.
• Upgradeability: green — the filter/inject is local to complete(); if pi-ai later emits instructions natively we can delete the hook in one place.
• No bloat: green — ~20 lines, no new deps, reuses existing toPiContext output (no helper refactor).
• Elegance: green — contract mirrors how openai-codex-responses already behaves upstream.

Test plan

• pnpm --filter @open-codesign/providers test -- --run (10 files, 134 tests, incl. 3 new cases: payload mutation, no-op when systemPrompt empty, no-op for anthropic wire)
• pnpm typecheck (10/10 tasks successful)
• pnpm lint (biome clean)
• Manual: call an openai-responses gateway that requires instructions and confirm 200 with system prompt honored

[github-actions]

Findings

• No high-confidence issues found in added/modified lines.

Summary

• Review mode: initial
• No issues identified in the latest diff.
• Residual risk/testing gap: behavior depends on pi-ai request payload shape for onPayload; repo tests cover unit-level mutation logic, but there is no in-repo integration test against a strict OpenAI-Responses gateway.

Testing

• Not run (automation)

open-codesign Bot

[github-actions]

Findings

• No high-confidence issues found in added/modified lines.

Summary

• Review mode: follow-up after new commits
• No issues identified in the latest diff.
• Not found in repo/docs: docs/VISION.md and docs/PRINCIPLES.md in this checkout; review used available repo context plus full PR diff.
• Residual risk/testing gap: behavior relies on pi-ai payload shape for onPayload; current coverage is unit-level and does not include an integration test against a strict OpenAI-Responses gateway.

Testing

• Not run (automation)

open-codesign Bot