config.get_value() re-reads and re-parses puppy.cfg from disk on every call — hot-path I/O churn + write race

Problem

code_puppy/config.py lines 252-256:

def get_value(key: str):
    config = configparser.ConfigParser()
    config.read(CONFIG_FILE)
    val = config.get(DEFAULT_SECTION, key, fallback=None)
    return val

Every one of the ~60 config getters funnels through this, and each call constructs a fresh ConfigParser and re-reads/parses the cfg file from disk. Several of these getters sit on hot paths:

• get_yolo_mode() and model_supports_setting() are called from make_model_settings() on every model construction.
• get_effective_model_settings() -> get_all_model_settings() does another full file read+parse, plus get_temperature() -> get_value() again -- 3+ file parses per model-settings resolution.
• The file's own comment at line ~200 admits get_global_model_name() "happens a LOT" -- which is why a _SESSION_MODEL cache was bolted on for that one key, but the underlying problem remains for everything else.

Why it matters

Dozens of redundant open()+parse syscalls per agent turn. The same pattern also creates a read-modify-write race in set_config_value()/set_model_name() (read file -> mutate -> rewrite whole file): two concurrent writers on different keys can silently drop one write.

Suggested fix

Cache the parsed config with mtime invalidation:

_cfg_cache: tuple[int, configparser.ConfigParser] | None = None

def _load_config() -> configparser.ConfigParser:
    global _cfg_cache
    try:
        mtime = os.stat(CONFIG_FILE).st_mtime_ns
    except OSError:
        mtime = -1
    if _cfg_cache is None or _cfg_cache[0] != mtime:
        cfg = configparser.ConfigParser()
        cfg.read(CONFIG_FILE)
        _cfg_cache = (mtime, cfg)
    return _cfg_cache[1]

def get_value(key: str) -> Optional[str]:
    return _load_config().get(DEFAULT_SECTION, key, fallback=None)

Writers (set_config_value, reset_value, clear_model_settings, set_model_name) should invalidate the cache after writing, ideally behind a single threading.Lock to also fix the read-modify-write race.

Filed by Zen Reviewer A (code-puppy-60635a)

Read code_puppy/config.py around get_value, set_config_value, reset_value, clear_model_settings, and set_model_name. Trace how reads and read-modify-write updates interact, then implement cache invalidation and synchronization so repeated getters avoid redundant parsing and concurrent writers do not lose updates.