fix(capabilities): eliminate prefix cache invalidation from in-place system prompt mutation

[Leoyzen]

Problem

MemoryCapability and SkillManagerCap modified SystemPromptPart.content in-place every turn via before_model_request, causing prefix cache misses on every model request. This increased token costs, latency, and wasted compute for both OpenAI (auto-prefix-cache) and Anthropic (prompt caching).

Closes #281.

Root Cause

Two capabilities used a shared _inject_into_system_prompt() utility that mutated SystemPromptPart.content directly in the message history:

# memory.py — _inject_into_system_prompt()
part.content = f"{part.content}\n\n{injected}"

Since the system prompt is part of the message history (not the instructions field), modifying it changes the prefix every turn → cache miss every turn.

Fix

Replace before_model_request in-place mutation with get_instructions returning callables. Pydantic-ai handles callables correctly: literal strings get dynamic=False (cacheable), callable results get dynamic=True (correct for dynamic content). No system prompt mutation occurs.

Changes

File	Change
memory.py	before_model_request → get_instructions returning async callable. Removed _inject_into_system_prompt + _is_model_request utilities.
skill_manager_cap.py	before_model_request → get_instructions returning [static_metadata, dynamic_callable]. Metadata is dynamic=False (cacheable); matched <skill_content> blocks are dynamic=True. Matcher uses RunContext.messages.
combined_toolset.py	Fixed get_instructions to preserve callables and sequences from children (previously silently dropped non-str returns).
dynamic_context.py	No change — message compaction is intentional behavior, not a cache invalidation bug.

How it works

Before (broken):

1. get_instructions() returns static metadata string
2. before_model_request() mutates SystemPromptPart.content in-place with dynamic content
3. System prompt changes every turn → prefix cache miss

After (fixed):

1. get_instructions() returns [static_str, async_callable]
2. Pydantic-ai calls the callable at run time with RunContext (has .messages)
3. Static string → InstructionPart(dynamic=False) → cacheable
4. Callable result → InstructionPart(dynamic=True) → correctly marked dynamic
5. No system prompt mutation → prefix cache preserved

Testing

• 48 tests pass (6 new tests for dynamic callable, matcher_fn, always_active, callable handling in CombinedToolsetCapability)
• ruff check clean
• ruff format --check clean
• mypy clean

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.