perf: avoid broad message prep deepcopies

[pedrommaiaa]

What does this PR do?

This reduces request-prep overhead by replacing broad message-history deepcopy() calls with targeted copy-on-write behavior.

The changed paths only mutate small parts of the outgoing request: provider-only tool call fields, prompt-cache markers, Qwen content normalization, and nested image payloads used during retry recovery. Copying the full message history in those cases is unnecessary, especially in long sessions with tool calls, reasoning metadata, and images.

Local request-prep benchmarks using Hermes-shaped payloads:

• chat cleanup, 1000-message history with dirty tool-call metadata: 9.699ms -> 0.963ms (10.1x faster)
• Anthropic prompt-cache prep, 201-message history with large content parts: 0.517ms -> 0.072ms (7.1x faster)
• Qwen prep, 502-message history with list content and image payloads: 1.185ms -> 0.126ms (9.4x faster)

This also tightens provider-switch safety:

• strict providers can strip Gemini-only metadata without mutating canonical history
• Gemini replay metadata survives strict-provider cleanup
• Gemma no longer receives stale Gemini extra_content
• Qwen cache_control does not leak into shared history
• nested image URLs are isolated before retry paths can mutate request payloads

Type of Change

• Bug fix
• Tests
• Refactor

Changes Made

• Use copy-on-write sanitization in agent/transports/chat_completions.py.
• Replace prompt-cache deepcopy() with targeted request-container copies in agent/prompt_caching.py.
• Replace Qwen OAuth deepcopy() with targeted normalization copies in plugins/model-providers/qwen-oauth/__init__.py.
• Add regression tests for provider switching, prompt caching, Qwen prep, Gemma/Gemini metadata handling, and nested image retry mutation.

How to Test

scripts/run_tests.sh tests/agent/test_prompt_caching.py tests/agent/transports/test_chat_completions.py tests/providers/test_provider_profiles.py tests/providers/test_profile_wiring.py tests/run_agent/test_provider_parity.py tests/run_agent/test_image_shrink_recovery.py -- -q

Result: 270 passed

scripts/run_tests.sh tests/run_agent/test_run_agent.py -- -q

Result: 367 passed

.venv/bin/ruff check agent/prompt_caching.py agent/transports/chat_completions.py plugins/model-providers/qwen-oauth/__init__.py tests/agent/test_prompt_caching.py tests/agent/transports/test_chat_completions.py tests/providers/test_provider_profiles.py tests/run_agent/test_provider_parity.py

Result: passed

[kshitijk4poor]

Merged via #61133 — your commit was cherry-picked onto current main with your authorship preserved (now 724ab90 on main). Thanks for the contribution! The copy-on-write conversions for ChatCompletionsTransport.convert_messages and QwenProfile.prepare_messages landed together with your mutation-safety tests.

Two pieces of this branch were intentionally scoped out (see #61133 for details): the agent/prompt_caching.py selective-copy (superseded by #57229, which rewrites that same function on current main) and the Gemma extra_content narrowing (reverts deliberate behavior from e8c3ac2 — worth its own PR with its own justification if you want to pursue it). Closing in favor of the merged salvage.