Enforce explicit chat-template kwarg allowlists

[hallerite]

Summary

Introduces an explicit, positive allowlist for the chat-template kwargs accepted by every model-specific renderer.

This PR is stacked on #103. Merge #123, then #103, then this PR.

Changes

• Every renderer config declares its chat-template fields explicitly.
• Class creation fails if a renderer-specific field is unclassified, classified twice, or references a nonexistent field.
• chat_template_kwargs rejects unknown, shared, and renderer-internal fields with an error that lists the allowed names.
• Renderer-internal options remain valid through the typed config.
• DefaultRenderer remains open to opaque Jinja kwargs while reserving its typed config fields.

Impact

This closes a validation hole where fields such as image_cache_max could be supplied through chat_template_kwargs even though they are not Jinja template inputs.

Verification

• 644 passed, 4 skipped in the config and chat-template kwarg parity suites
• Ruff lint and formatting pass
• Ty completes with the existing repository warning baseline

Note

Enforce explicit chat-template kwarg allowlists and add InklingRenderer with multimodal support

• Adds _template_fields and _allow_opaque_template_kwargs to BaseRendererConfig, requiring every renderer config subclass to explicitly classify its fields; unclassified fields raise TypeError at import time.
• Updates _merge_chat_template_kwargs in renderers/base.py to validate chat_template_kwargs against each config's declared allowlist, raising ValueError for unknown or renderer-internal keys.
• Introduces InklingRenderer and InklingRendererConfig in renderers/inkling.py supporting image and audio multimodal content, tool declaration/invocation, reasoning_effort control, thinking retention, and bridge_to_next_turn.
• Adds parse_inkling in renderers/parsing.py to decode Inkling model outputs into structured ParsedResponse objects including reasoning content and tool calls.
• Risk: passing previously-accepted but undeclared chat_template_kwargs to typed renderers now raises ValueError instead of silently merging.

^{Macroscope summarized 8d5965a.}

---

Note

Medium Risk
Changes how invalid chat_template_kwargs fail (often ValueError at merge time instead of pydantic ValidationError), which may break callers that relied on passing internal fields through kwargs; behavior for valid configs is stricter and more explicit.

Overview
Enforces explicit allowlists for chat_template_kwargs so renderer-internal options (e.g. image_cache_max, tool_parser) can’t be smuggled in as if they were Jinja template knobs.

Each renderer config now declares _template_fields and _internal_fields; template_field_names() returns only the template set. __pydantic_init_subclass__ fails at class definition time if any field is missing, duplicated, or unknown in that split.

_merge_chat_template_kwargs checks incoming keys against that allowlist and raises ValueError with allowed names; renderer-only fields must go on the typed config. DefaultRendererConfig keeps _allow_opaque_template_kwargs for arbitrary Jinja extras but still blocks typed fields like tool_parser from the kwargs map.

README and docs/renderer-config.md describe the new contract; tests cover allowlist validation, internal-field rejection, and default-renderer open-ended vs reserved behavior.

^{Reviewed by Cursor Bugbot for commit 635d351. Bugbot is set up for automated code reviews on this repo. Configure here.}

[macroscopeapp]

Approvability

Verdict: Approved 82bff80

This PR adds explicit kwarg validation and field classification to renderer configs - a guardrails/type-safety improvement that makes the system more restrictive rather than enabling new behavior. Changes are self-contained, well-tested, and the author has prior contributions to these files.

No code changes detected at 635d351. Prior analysis still applies.

^{You can customize Macroscope's approvability policy. Learn more.}