[codex] Fix open bug regressions and core build readiness

[mimeding]

Business rationale

This PR clears several high-impact harness reliability bugs that make osaurus feel untrustworthy when users connect real providers, stream reasoning/tool calls, download models, or run sandboxed chat workflows. It advances the multi-provider, agent, sandbox, and file-safety pillars by making provider payloads round-trip correctly, keeping tests away from production chat-history storage, and making local capacity checks reflect what macOS actually reports.

Coding rationale

The fixes stay narrow and preserve existing public shapes: provider wire models keep optional booleans where omission and false have different semantics, streamed reasoning/tool-call state is preserved instead of being flattened into visible chat text, and free-space checks are routed through URL volume resource keys rather than legacy filesystem attributes. The rebase cleanup also makes the touched files pass both SwiftFormat and SwiftLint without changing behavior; multiline optional bindings were split into small locals, and SwiftUI trace calls keep explicit discard bindings because ViewBuilder treats plain assignment as an invalid view expression.

What changed

• Fix MLX/Jinja tool-spec serialization by dropping null placeholders before tokenizer conversion.
• Preserve DeepSeek/OpenAI-compatible reasoning_content across streamed and non-stream tool-call turns so follow-up requests can replay it.
• Query storage capacity through URL volume resource keys and reuse that path for model-download free-space checks.
• Accept OpenAI Responses shorthand input items where message objects omit optional type: "message".
• Update FluidAudio PocketTTS initialization for the required language parameter and align FluidAudio lockfiles so OsaurusCore compiles against the current resolved dependency set.
• Keep core tests off production chat-history storage when no test root is configured.
• Harden sandbox context-budget regression coverage.
• Rebase on current origin/main and make all touched Swift files pass strict formatting/lint.

Fixes #980
Fixes #959
Fixes #964
Fixes #662

Validation

• swift build --package-path Packages/OsaurusCore
• swift build --package-path Packages/OsaurusCore -c release
• swift test --package-path Packages/OsaurusCore
• xcrun swift-format lint --strict <touched Swift files>
• swiftlint lint --strict <touched Swift files>
• git diff --check origin/main...HEAD
• git diff --check

Non-scope

• No changes to PR [codex] Use lightweight context for Foundation #979's Foundation context trimming.
• No changes to PR [codex] Preserve system prompts for Gemma templates #992's Gemma/system-prompt behavior.
• No new provider integrations or model-profile policy changes.

Residual risks

• SwiftPM still reports existing dependency/resource warnings from current main and transitive packages; this PR does not broaden scope to clean unrelated warnings.
• GitHub Actions must confirm the rebased branch in CI after the force-with-lease update.

[mimeding]

CI note: shellcheck, swiftlint, test-cli, and Release Drafter passed. test-core failed before running Osaurus tests with the known Xcode/SwiftPM C-module dependency resolution issue in EventSource dependencies (CAsyncHTTPClient, CNIOLLHTTP, CNIOExtrasZlib, CNIOPosix, _NumericsShims). Local validation on the rebased branch passed with swift test --package-path Packages/OsaurusCore (1174 tests). I attempted gh run rerun --failed, but this account does not have permission to rerun upstream Actions.

[mimeding]

@tpae Please ask someone with repo permissions to rerun test-core on PR #985.

[tpae]

can we remove this

[mimeding]

Author-side cleanup complete.

Verification:

• GitHub checks are green: test-core, test-cli, swiftlint, shellcheck, update_release_draft, and pr-clean-gate.
• scripts/ci/check-pr-clean.sh osaurus-ai/osaurus 985 reports clean.
• Local verification from the clean worktree passed for swift build, targeted OsaurusCore tests, and shellcheck on the clean-PR gate.

Root cause addressed: stale DerivedData restore on PRs plus stale FluidAudio locks. This branch now keeps PR test-core on the cold DerivedData path and uses FluidAudio 0.14.3.

[mimeding]

@tpae the line/rule you flagged has been removed, and #985 is now passing clean on the latest head. Good to commit from my side.

[mimeding]

@tpae The line was removed, and this is now good to commit. Latest checks are green: test-core, test-cli, swiftlint, shellcheck, and Release Drafter.

[tpae]

Hey @mimeding i apologize for not able to merge this, wanted to add DeepSeek as official provider since reasoning_content only applies to DeepSeek at the moment.

I added official DeepSeek provider here: #1012