perf(agent): AIAgent hot-path salvage — prompt-cache copy, reasoning-timeout precompute, lazy compressor init

[kshitijk4poor]

Summary

Combined salvage of three verified, non-conflicting AIAgent hot-path
performance fixes, rebased onto current main with original authorship
preserved.

Supersedes #57046, #56081, #38991 — these can be closed in favor of this PR.

What's included

perf(caching): selective copy instead of deepcopy entire history — @kohoj (#57046)

apply_anthropic_cache_control deep-copied the entire message history on
every Anthropic-wire API call (agent/prompt_caching.py), from the
run_conversation hot path. It now shallow-copies the list and deep-copies
only the ≤4 messages that receive a cache_control marker. Merged with main's
newer static_system_prefix two-marker system split (the system message is
deep-copied before _apply_system_cache_markers).

• Verified: output byte-identical to a reference full-deepcopy implementation
  across native_anthropic × TTL × static-prefix combos; input list never mutated.
• Original benchmark (401-message / ~3 MB history): 0.78 ms → 0.02 ms/call (~39×),
  paid every turn.

perf: pre-compute sorted reasoning timeout floors — @Stoltemberg (#56081)

_match_any() re-sorted the floors table and lazily compiled regexes into a
mutable dict on every call (agent/reasoning_timeouts.py). Sorted, pre-compiled
(slug, floor, pattern) tuples are now built once at import — faster and
thread-safe. Verified: 162 probe model names resolve identically to main.

perf(agent): defer synchronous httpx probe out of AIAgent.init — @rodboev (#38991, closes #32221)

ContextCompressor.__init__ resolved the model context length eagerly via
get_model_context_length(), which can issue a synchronous /models HTTP
probe — blocking agent construction. Resolution is deferred to first access via
lazy properties (with setters for direct assignment). Main's newer
small-context threshold floor (75% under 512K) and absolute threshold_tokens
cap are applied at first resolution / first threshold derivation, so behavior
after resolution is identical to eager init.

fix(compression): keep ContextCompressor init non-blocking when quiet_mode=False — @kshitijk4poor

The "Context compressor initialized" log line (quiet_mode=False, interactive
CLI) reads context_length/threshold_tokens/tail_token_budget — property
reads that resolve the deferred value, so #38991's deferral silently didn't
fire on the non-quiet path (its tests only exercised quiet_mode=True). The line
is now emitted once, on first resolution. Regression test asserts no probe in
__init__ for both quiet modes and exactly one init log on first access.

fix(compression): copy-on-write in image-shrink recovery — @kshitijk4poor

Found by the Hermes-specific review pass of the selective-copy change:
un-marked messages on the decorated per-call list share nested content parts
with the persistent conversation history. try_shrink_image_parts_in_messages
wrote re-encoded images INTO those aliased part/source dicts, so an
image-too-large retry on an Anthropic route would replace the original image
bytes in agent.messages and persist the degraded copy. Now rebuilds the
content list copy-on-write and reassigns msg["content"] (top-level write on
the per-call copy). Two regression tests simulate the aliasing; both fail
against the in-place implementation (mutation-verified).

Test plan

• Targeted suites: tests/agent/test_context_compressor.py,
  test_compress_focus.py, test_prompt_caching.py,
  test_reasoning_stale_timeout_floor.py — 320 passed
• Image-shrink + routing + caching suites — 182 passed
• Broad baseline diff (tests/agent/ + 9 run_agent compression files +
  context_switch_guard, ~7,200 tests) run on this branch AND clean
  upstream/main: identical failure sets (182 pre-existing environmental
  failures on both) — zero PR-caused regressions
• E2E with real imports, temp HERMES_HOME:
  prompt_caching byte-identical vs reference deepcopy impl, input never
  mutated (8 combos); 162 reasoning-timeout probes identical to old impl;
  no probe in __init__ (both quiet modes), floor applied on first
  resolution, single probe
• ruff clean on all touched files

Attribution

contributors/emails/tutors1997@outlook.com → Stoltemberg added (kohoj and
rodboev already mapped). Commit authorship preserved in git history;
contributor_audit --strict passes.

[alt-glitch]

This was generated by AI during triage.

Salvage/superset of three OPEN perf PRs (Supersedes #57046, #56081, #38991) cherry-picked onto main — related, not duplicates (salvage keeps credit). Also related to the sibling hot-path bundle #54710. Note for review: the prompt_caching.py change is a copy-mechanism optimization only — the cache_control markers are applied at identical indices with an identical breakpoint budget, so the prompt-cache invariant is unchanged (no caching risk).