fix(core): redact image bytes from debug dumps by default

[bug-ops]

Summary

• debug_dump/mod.rs serialized outbound LLM requests verbatim, including MessagePart::Image base64 payloads, in both json and raw dump formats. The raw format's dominant path clones the provider's native wire payload directly, so Claude/OpenAI/Gemini/Ollama vision requests all leaked full image bytes to disk whenever debug dumps were enabled ([debug] enabled = true, opt-in feature).
• All known image-payload shapes (internal typed form, Claude, OpenAI, Gemini, Ollama) are now redacted to a <redacted image: {mime_type}, {n} bytes, blake3:{prefix}> marker via a single generic recursive JSON-value redactor applied to both dump formats.
• Added an opt-in [debug] include_raw_images config flag (default false) to restore full raw byte output for wire-payload debugging; enabling it now emits a tracing::warn! at construction time so the tradeoff is visible in logs.

Notes

• Scope is intentionally bounded to MessagePart::Image in outbound requests. Other sensitive text content in messages (PII, secrets) is unaffected by this change, and dump_tool_output/dump_response have a related but out-of-scope residual leak vector (raw tool output text, e.g. a vision tool returning base64 as text) — filing a follow-up issue for that.
• Adversarial review (impl-critic) flagged that the initial test suite was self-referential (hand-authored JSON matching the redactor's own shape assumptions, fail-open risk if a provider's wire shape drifts). Addressed with a regression test that drives the actual ClaudeProvider/OpenAiProvider/GeminiProvider/OllamaProvider serializers from zeph-llm with an image-bearing message and asserts no raw base64 survives the real dump_request/raw_dump path.
• OpenAI's Responses API input_image string shape (a known future leak vector) is not wired up anywhere in this codebase today — no dead-code test added; documented in the playbook.

Closes #6306

Test plan

• cargo +nightly fmt --check — clean
• cargo clippy --profile ci --workspace --all-targets --features "desktop,ide,server,chat,pdf,scheduler,testing" -- -D warnings — clean
• cargo nextest run --config-file .github/nextest.toml --workspace --features "desktop,ide,server,chat,pdf,scheduler" --lib --bins — 13802 passed, 0 failed, 35 skipped
• RUSTFLAGS="-D warnings" RUSTDOCFLAGS="--deny rustdoc::broken_intra_doc_links" cargo doc --no-deps --workspace --features "desktop,ide,server,chat,pdf,scheduler" — clean
• gitleaks protect --staged --no-banner --redact — no leaks
• New regression test drives real provider serializers (Claude/OpenAI/Gemini/Ollama) and asserts redaction on actual wire-format JSON, not hand-authored fixtures
• Live-session verification against a real vision-capable provider — not yet exercised with a live API key; tracked as Untested in .local/testing/coverage-status.md