fix(openai-compat): recover from malformed tool calls

[xjdr-noumena]

Summary

Clean malformed OpenAI-compatible tool-call recovery for #48 and the useful part of #52.

• adds a pre-send history sanitizer that strips malformed assistant tool_use blocks with blank, missing, or non-string names
• removes matching tool_result blocks so bad tool-call history does not poison every later OpenAI-compatible request
• adds schema-confusion detection for likely sibling-tool argument swaps across Bash, Grep, Glob, Read, WebFetch, and Write
• appends targeted model-facing hints to the existing Zod validation error instead of silently rerouting tool calls
• emits structured telemetry for rejected schema leaks via ncode_schema_leak_rejected
• adds focused tests for sanitizer behavior, schema-confusion hinting, and telemetry shape

Scope intentionally excluded from this PR:

• retry-loop breaking after repeated failures
• user-visible UI warnings
• resume-time poisoned-history guardrails
• schema-router/auto-reroute behavior
• CI, dependency, release, or unrelated cleanup changes from fix: sanitize empty-name tool calls to prevent session-poisoning 400 loops #52

This should be treated as partial #48 coverage plus the real #52 history-sanitizer fix.

Validation

Focused recovery tests:

/home/xjdr/.bun/bin/bun test \
  src/services/api/openAICompatInferenceClient.test.ts \
  src/services/tools/schemaConfusionHint.test.ts \
  src/services/tools/schemaLeakTelemetry.test.ts

Result: 93 pass, 0 fail.

Contracts:

PATH="$HOME/.bun/bin:$PATH" timeout 120s /home/xjdr/.bun/bin/bun run test:contracts

Result: 238 pass, 0 fail.

Package smoke:

PATH="$HOME/.bun/bin:$PATH" timeout 180s /home/xjdr/.bun/bin/bun run test:package-smoke

Result: passed. Security audit OK; native probe OK with documented sharp-fallback image mode.

Diff check:

git diff --cached --check

Result: passed before commit.