perf(api-server): normalize chat content linearly

[teknium1]

Summary

API-server text content normalization now tracks accumulated length with a running counter instead of rescanning every collected part on each loop iteration.

Changes

• gateway/platforms/api_server.py: replace repeated sum(len(part) for part in parts) with total_len updates for strings, text parts, and nested lists.
• tests/gateway/test_api_server_normalize.py: add a regression test that monkeypatches the module-level sum name and proves normalization no longer calls it in the large-list path.

Validation

Check	Result
RED test before implementation	failed: sum_calls == 1000
Focused regression	1 passed in 0.15s
Microbench	1000 parts median 0.1481ms, min 0.1461ms
Targeted API server tests	175 passed in 3.9s
Syntax/diff checks	py_compile + git diff --check passed

Notes

• Output behavior is preserved: same newline joins, same text-part handling, same final truncation limit.
• The untracked .plans/speed-optimization-audit.md remains intentionally excluded from this PR.

Infographic

[github-actions]

🔎 Lint report: hermes/api-server-normalize-linear vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 10895 on HEAD, 10895 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 5726 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.