fix(memory): avoid rejected mid-conversation system injection on Claude when preceding turn isn't a tool result

[diegosouzapw]

Summary

Closes #11290.

Claude Opus 5 rejects OmniRoute's #3890 cache-safe mid-array system-message
splice with HTTP 400 in one specific shape: when the assistant turn
immediately before the splice point is plain text rather than a server-side
tool result (*_tool_result content block). injectMemory()
(src/lib/memory/injection.ts) always spliced the memory-context system
message right before the last user turn whenever cacheSafe was set and the
provider was not in the systemMessageMustBeFirst() strict set — and
anthropic/claude were never in that set, so every Claude cache-safe
request hit the same mid-array shape regardless of what preceded it.

Design decision

Two options were on the table:

1. Add claude/anthropic to BUILTIN_PROVIDERS_SYSTEM_MUST_BE_FIRST
   (systemMessageMustBeFirst()), forcing every Claude request onto the
   leading-system-message path.
2. Narrow the condition so the mid-array splice is only skipped for the
   specific requests that trigger the 400.

Option 1 is the simplest change but reverts the #3890 prompt-cache-hit
optimization for every Claude request that uses prompt caching, including
the (presumably majority) requests where the preceding turn already ends in a
tool result and the splice works fine — a real cost regression for no gain
in those cases.

This PR takes option 2: injectMemory() gained a narrower, additive check
that only applies when cacheSafe is on, the provider is in the Claude
family (claude, anthropic, isAnthropicCompatibleProvider() — which also
covers isClaudeCodeCompatibleProvider()'s narrower prefix), and the message
immediately before the splice point is not an assistant turn ending in a
*_tool_result content block. In that case it falls back to the existing
injectSystemFirst() leading-message placement (merging into the request's
system message, exactly like the strict-provider path already does) instead
of splicing. The mid-array splice is preserved everywhere it is safe: for all
non-Claude providers unconditionally, and for Claude providers whose
preceding turn does end in a server tool result.

Changes

• src/lib/memory/injection.ts: added isClaudeFamilyProvider() and
  endsWithServerToolResult() helpers, and a new gate in injectMemory()
  applied after the existing systemMessageMustBeFirst() check.
• src/lib/memory/__tests__/injection.test.ts: TDD regression test proving
  the fallback (RED before the fix, GREEN after — verified locally by
  reverting injection.ts and re-running the new test in isolation).
• tests/unit/memory-cache-safe-injection.test.ts: the 3 pre-existing
  anthropic-provider cache-safe tests encoded exactly the now-fixed buggy
  shape (mid-array splice after a plain-text assistant turn), so they were
  switched to a non-Claude-family provider (openai) to keep testing the
  original #3890 generic mechanism without regressing it. Added a new
  describe block with 4 tests covering: the plain-text fallback, the
  preserved splice when the preceding turn is a server tool result, the
  Claude-Code-compatible passthrough provider id, and confirmation that
  non-Claude providers are unaffected.

Test plan

• node --import tsx/esm --test tests/unit/memory-cache-safe-injection.test.ts tests/unit/memory-glm-injection.test.ts tests/unit/chatcore-memory-skills-injection.test.ts — all green (39 tests)
• npx vitest run src/lib/memory/__tests__/injection.test.ts — all green (23 tests)
• Verified RED→GREEN: reverted injection.ts to the pre-fix version, re-ran the new #11290 vitest test — it failed as expected (mid-array splice still occurred), then restored the fix — it passed.
• npx eslint on the 3 changed files — clean (no new no-explicit-any beyond the pre-existing, suppressed count for memory-cache-safe-injection.test.ts).
• tsc --pretty false -p tsconfig.typecheck-core.json — clean, no errors.
• npm run check:cycles — no cycles introduced by the new src/shared/constants/providers import.

⚠️ base-red inherited: #9985