fix: strip NSNull from chat templates before Jinja render (#168)

[CodeAndCanvas728]

Fixes #168

Root cause

The crash was not in the model's chat template. It happens before render:

1. AnyCodable decodes JSON null → NSNull (common in opencode/zod tool schemas: "default": null, etc.)
2. Tokenizer.applyChatTemplate converts tools with Jinja.Value.init(any:), which has no NSNull case
3. swift-jinja 2.3.5 reported a misleading Optional<Any> (type(of:) on the Any? parameter); 2.5.1 correctly reports NSNull but still throws
4. The error was wrapped as MalformedChatTemplate ("defect in the model's chat_template.jinja... not in the request")
5. The fix: name a malformed chat template instead of failing every request #142 startup probe used tools: nil, so it never exercised this path

Verified: without this fix, a tools request with default: null against Ministral-8B returns 500 Cannot convert value of type NSNull to Jinja Value. With the fix, the same request (plus X-SwiftLM-Prefill-Progress) streams cleanly.

Changes

A. Root fix — sanitizeForJinja

Recursively drops NSNull (and unwraps nested optionals) from messages, tools, and additionalContext inside TransformersTokenizerBridge.applyChatTemplate so every caller is covered.

B. Startup probe — tools case (warn-only)

Second render with a minimal non-empty tools array (parameters include a null). On failure: loud [SwiftLM] ⚠️ warning; server continues (plain chat / /v1/completions still work).

C. Heartbeat payload

ssePrefillChunk now includes "choices": [] so opencode's strict ChatCompletionChunk union accepts the named prefill_progress event (named event alone was not enough — opencode validates every data: line).

D. Dependency

swift-jinja 2.3.5 → 2.5.1 (accurate conversion errors, Optional/BinaryInteger bridging). Alone this does not fix NSNull — A is required.

E. Tests

• JinjaSanitizerTests (6 cases: nested objects, arrays, optionals, scalars)
• Updated ServerSSETests to expect empty choices array
• tests/test-opencode.sh Test 3: tools + prefill-progress combined (closes the Test 1 / Test 2 coverage gap called out in Chat template crash (Optional<Any> Jinja error) only triggers when tools are present — startup probe (#142) misses it #168 / feat: OpenAI-compatible streaming hardening (prefill heartbeat + OpenCode e2e CI) #75)

Verification

• Full SwiftLMTests suite: 174/174 pass
• Release build succeeds
• Baseline (this commit stashed): tools+null → 500 with NSNull error
• With fix: tools+null+heartbeat against mlx-community/Ministral-8B-Instruct-2410-4bit → stream accepted
• tests/test-opencode.sh full run (script pins gemma-4-e4b, not available in this environment)

Notes

• Issue title/message still blames the checkpoint template; after this fix the Tiel template itself is fine for null-bearing tool schemas. A genuine tools-branch parse error would now warn at startup instead of failing every request silently.
• Pre-existing SwiftBuddyTests VLM/Audio failures (missing binary at .build/arm64-apple-macosx/debug/SwiftLM) are unrelated.

🤖 Generated with Claude Code

[CodeAndCanvas728]

Closing in favor of #170, which contains this commit (ff0c28d) plus the Xcode 26.3 compile fix for (any Sendable)?, per maintainer review.