candor_classify/

lib.rs

//! candor-classify — the curated effect classifier (crate+path -> effect), extracted to a STABLE
//! crate so both the nightly `rustc_private` lint AND a stable backend share ONE source of truth
//! (no drift). Pure string logic; no rustc internals. The effect vocabulary lives in candor-report.

use candor_report::EFFECTS;

/// The canonical CANDOR_POLICY DSL parser (SPEC §6.2), shared by the nightly gate and candor-query.
pub mod policy;

/// Project-supplied rules, consulted only when the built-in `classify` returns None.
pub fn classify_extra(
    crate_name: &str,
    path: &str,
    extra: &[(&'static str, bool, String)],
) -> Option<&'static str> {
    for (eff, is_crate, prefix) in extra {
        let hit = if *is_crate { crate_name.starts_with(prefix.as_str()) } else { path.starts_with(prefix.as_str()) };
        if hit {
            return Some(eff);
        }
    }
    None
}

/// The exact third-party crates `classify` has effect rules for, and the crate-name
/// PREFIXES it recognizes. This is the single source of truth for "what candor knows":
/// it is emitted beside the JSON report (`<prefix>.calibrated.json`) so the Claude Code
/// receipt's coverage check reads candor's real coverage instead of a hand-copied list.
/// Keep in lockstep with `classify` below — the `db_crates_are_calibrated` and
/// `calibrated_crates_are_live` tests (in this crate's `tests` module) enforce both directions.
pub const CALIBRATED_CRATES: [&str; 53] = [
    // network (aws_config resolves credentials over the network on `.load()`;
    // git2 remote ops — fetch/push/connect — contact the network; async_net is smol's net layer;
    // pnet is raw L2/L3 packet capture)
    "reqwest", "isahc", "ureq", "curl", "aws_config", "git2", "tokio_tcp", "tokio_udp", "async_net",
    "async_nats", "lapin", "lettre", "tungstenite", "elasticsearch", "tonic", "rdkafka", "pnet",
    // directory traversal (ignore = gitignore-aware walker, powers ripgrep/fd; its walk executors are Fs)
    // + filesystem watching (notify = inotify/FSEvents/kqueue wrapper; powers watchexec/cargo-watch)
    "ignore", "notify",
    // database (see DB_CRATES in classify)
    "sqlx", "rusqlite", "postgres", "tokio_postgres", "diesel", "redis", "mongodb",
    "mysql", "mysql_async", "sea_orm", "deadpool_postgres",
    // filesystem (async_fs = smol; fs_err = std::fs wrapper; tempfile; glob) / entropy /
    // subprocess (async_process = smol; duct) / env (dotenvy/dotenv) / clock (time) / log / clipboard
    "memmap2", "fs_err", "async_fs", "tempfile", "glob",
    "rand", "getrandom", "fastrand",
    "portable_pty", "async_process", "duct",
    "dotenvy", "dotenv",
    "chrono", "time", "tracing", "log", "arboard",
    // compiler diagnostic emission (a dylint lint's output) — see the Log rules in classify
    "rustc_lint", "rustc_errors",
    // raw syscalls via FFI — the syscall-name table that lights up the FFI-thin tier (nix is routed
    // through the same table by leaf name, so a consumer of nix is covered without nix's own source)
    "libc", "nix", "rustix",
];

pub const CALIBRATED_PREFIXES: [&str; 3] = ["aws_sdk_", "aws_smithy", "cap_"];

/// Crates `classify` matches by PATH prefix rather than crate-name equality (their effectful modules
/// are recognised, e.g. `tokio::net::`/`async_std::fs::`/`mio::net::`), so they're absent from
/// `CALIBRATED_CRATES` (which the liveness test probes by crate name). The coverage check must still
/// treat them as *covered* — otherwise it would mislabel the most common async crates as blind spots.
pub const PATH_CALIBRATED_CRATES: [&str; 3] = ["tokio", "async_std", "mio"];

/// Representative path tails (each appended to a crate name) that the `calibrated_crates_are_live`
/// liveness test probes: at least one must match for every `CALIBRATED_CRATES` entry, else the entry is
/// dead. Exported as ONE source of truth because the nightly lint crate (`src/lib.rs`) runs the SAME
/// liveness test — when the two probe lists were duplicated they drifted, and a rule keyed on a
/// distinctive tail (pnet `::datalink::channel`, ignore `::WalkBuilder::build_parallel`, notify
/// `::RecommendedWatcher::new`) added to only one list silently broke the other crate's `cargo test`.
pub const CALIBRATION_PROBE_TAILS: &[&str] = &[
    "::X::send", "::X::execute", "::X::call", "::X::query", "::X::fetch_one", "::Remote::fetch",
    "::datalink::channel", "::WalkBuilder::build_parallel", "::RecommendedWatcher::new",
    "::X::connect", "::Utc::now", "::X::load", "::__private_api::log", "::tempfile", "::glob",
    "::X::run", "::dotenv", "::random", "::emit", "::X::emit_span_lint", "::X::anything",
];

/// Database client crates whose execution verbs are I/O (see the DB branch in `classify`).
/// Module-level so `db_crates_are_calibrated` can enforce `DB_CRATES ⊆ CALIBRATED_CRATES`.
pub const DB_CRATES: [&str; 11] = [
    "sqlx", "rusqlite", "postgres", "tokio_postgres", "diesel", "redis", "mongodb",
    "mysql", "mysql_async", "sea_orm", "deadpool_postgres",
];

/// Classify a resolved callee by the crate it belongs to and its full path.
pub fn classify(crate_name: &str, path: &str) -> Option<&'static str> {
    if crate_name.starts_with("aws_sdk_") || crate_name.starts_with("aws_smithy") {
        // Only request dispatch is network I/O; builder setters/accessors are pure.
        if path.ends_with("::send") || path.ends_with("::send_with") {
            return Some("Net");
        }
        return None;
    }
    // aws-config resolves credentials/region on `.load()` — it reaches the IMDS metadata
    // endpoint / STS over the network (and reads ~/.aws + env). Builders (`defaults()`,
    // `SdkConfig::builder()`, `BehaviorVersion::latest()`) are pure; the `load` is the I/O.
    // (Found hardening on a real app, ebman: `builder.load().await` was classified pure.)
    if crate_name == "aws_config" {
        if path.ends_with("::load") || path.ends_with("::load_defaults") {
            return Some("Net");
        }
        return None;
    }
    // git2 (libgit2 FFI): remote operations contact the network; everything else is local
    // to the .git directory. Match the remote verbs precisely — NOT bare `::clone`, which is
    // the `Clone`-trait dup of a `Remote` handle (pure), not `Repository::clone`. (Found
    // hardening on gitui: `remote.fetch`/`remote.push` were classified network-free — a git
    // client reporting it makes no network calls.)
    if crate_name == "git2" {
        if path.ends_with("::fetch")
            || path.ends_with("::push")
            || path.ends_with("::download")
            || path.ends_with("::connect")
            || path.ends_with("::connect_auth")
            || path.ends_with("::ls")
            || path.ends_with("::upload")
        {
            return Some("Net");
        }
        return None;
    }
    // libc — raw syscalls via FFI. The FFI-thin tier (nix, and the syscall layer beneath rusqlite/git2)
    // is invisible to a name classifier unless we model libc directly: a 35-crate calibration
    // (eval/calibration) showed nix reporting ZERO library effects because every wrapper bottoms out in
    // an unrecognised `libc::*` call. Classify by syscall name, but ONLY the UNAMBIGUOUS ones — the
    // socket family is Net, path/dir syscalls are Fs, spawn/exec/wait is Exec, SysV/pipe IPC is Ipc,
    // env/clock/entropy each their own. We deliberately SKIP the generic file-descriptor ops
    // (read/write/close/lseek/dup/fcntl/ioctl/poll/select/epoll*/mmap): they operate on ANY fd — file,
    // socket, or pipe — so a fixed label would mis-categorise as often as it helps. An honest
    // no-classify (under-report) beats emitting the WRONG effect. Pure conversions (htons/inet_pton/
    // gmtime) are also skipped.
    //
    // `nix` (the idiomatic SAFE libc wrapper, in ~every Rust systems/CLI crate) is routed through the
    // SAME table: its functions keep the syscall leaf name (`nix::fcntl::open`, `nix::sys::socket::connect`,
    // `nix::unistd::execvp`). Without this, a CONSUMER of nix analysed without nix's own source (the
    // stable scanner, single-crate) sees `nix::*` cross-crate and under-reports — serialport-rs opens its
    // device via `nix::fcntl::open` and reported ZERO Fs. The nightly lint reaches `libc::*` THROUGH nix's
    // body; this gives the scanner the same coverage directly. (Found sweeping serialport-rs.)
    // `rustix` is the same shape as nix but does RAW syscalls (no libc underneath), so its functions MUST
    // be classified directly. Its leaf names are the syscall names too (`rustix::time::clock_settime`,
    // `rustix::fs::mkfifoat`/`symlink`/`stat`, `rustix::net::connect`) — route it through the same table.
    // The rustix-specific `*at`/variant leaves it doesn't share with libc just under-report (the safe
    // direction). VALIDATED, not speculative: coreutils' `date` reads/sets the clock via
    // `rustix::time::clock_getres`/`clock_settime` and reported Clock=0; the file I/O that goes through
    // std::fs was already correct, which is why only the rustix-only effects (Clock/Ipc) were missing.
    if crate_name == "libc" || crate_name == "nix" || crate_name == "rustix" {
        let f = path.rsplit("::").next().unwrap_or(path);
        // path / directory / metadata syscalls (incl. *64 and *at variants)
        const FS: &[&str] = &[
            "open", "open64", "openat", "openat2", "creat", "creat64", "stat", "stat64", "lstat",
            "lstat64", "fstatat", "fstatat64", "newfstatat", "statx", "access", "faccessat",
            "faccessat2", "mkdir", "mkdirat", "rmdir", "unlink", "unlinkat", "rename", "renameat",
            "renameat2", "link", "linkat", "symlink", "symlinkat", "readlink", "readlinkat", "chmod",
            "fchmodat", "chown", "lchown", "fchownat", "truncate", "truncate64", "ftruncate",
            "ftruncate64", "opendir", "fdopendir", "readdir", "readdir64", "readdir_r", "closedir",
            "rewinddir", "seekdir", "telldir", "scandir", "mkstemp", "mkstemps", "mkostemp", "mkdtemp",
            "mknod", "mknodat", "chdir", "fchdir", "getcwd", "get_current_dir_name", "chroot",
            "pivot_root", "statfs", "statfs64", "fstatfs", "fstatfs64", "statvfs", "fstatvfs", "mount",
            "umount", "umount2", "fsync", "fdatasync", "sync", "syncfs", "sync_file_range", "fallocate",
            "posix_fallocate", "posix_fadvise", "sendfile", "sendfile64", "copy_file_range", "flock",
            "getdents", "getdents64", "utime", "utimes", "lutimes", "futimens", "utimensat", "futimesat",
            "realpath",
        ];
        // socket family — these operate only on sockets, so Net is unambiguous (AF_UNIX domain isn't
        // visible at the call, so a Unix socket reads as Net rather than Ipc; acceptable over-general).
        const NET: &[&str] = &[
            "socket", "setsockopt", "getsockopt", "bind", "listen", "accept", "accept4", "connect",
            "shutdown", "send", "sendto", "sendmsg", "sendmmsg", "recv", "recvfrom", "recvmsg",
            "recvmmsg", "getpeername", "getsockname", "getaddrinfo", "freeaddrinfo", "getnameinfo",
        ];
        // process creation / replacement / reaping
        const EXEC: &[&str] = &[
            "fork", "vfork", "clone", "clone3", "execl", "execlp", "execle", "execv", "execvp",
            "execvpe", "execve", "execveat", "fexecve", "posix_spawn", "posix_spawnp", "system",
            "popen", "pclose", "wait", "waitpid", "wait3", "wait4", "waitid",
        ];
        // pipes / FIFOs / SysV + POSIX message queues, semaphores, shared memory; socketpair (AF_UNIX)
        const IPC: &[&str] = &[
            "pipe", "pipe2", "mkfifo", "mkfifoat", "socketpair", "msgget", "msgsnd", "msgrcv", "msgctl",
            "semget", "semop", "semtimedop", "semctl", "shmget", "shmat", "shmdt", "shmctl", "mq_open",
            "mq_send", "mq_receive", "mq_timedsend", "mq_timedreceive", "mq_close", "mq_unlink",
        ];
        const ENV: &[&str] = &["getenv", "secure_getenv", "setenv", "putenv", "unsetenv", "clearenv"];
        const CLOCK: &[&str] = &[
            "time", "gettimeofday", "clock_gettime", "clock_getres", "nanosleep", "clock_nanosleep",
            // SETTING the system clock is a clock effect too (was unclassified — found on coreutils `date`,
            // which sets it via `clock_settime`).
            "clock_settime", "settimeofday", "stime", "adjtime", "adjtimex", "clock_adjtime",
        ];
        const RAND: &[&str] = &["getrandom", "getentropy", "arc4random", "arc4random_buf", "arc4random_uniform"];
        if FS.contains(&f) {
            return Some("Fs");
        }
        if NET.contains(&f) {
            return Some("Net");
        }
        if EXEC.contains(&f) {
            return Some("Exec");
        }
        if IPC.contains(&f) {
            return Some("Ipc");
        }
        if ENV.contains(&f) {
            return Some("Env");
        }
        if CLOCK.contains(&f) {
            return Some("Clock");
        }
        if RAND.contains(&f) {
            return Some("Rand");
        }
        return None;
    }
    // C-library FFI bindings: libsqlite3 (under rusqlite) and libgit2 (under git2). Like the libc tier,
    // these crates are thin Rust over a C library, so their real I/O is invisible until the C entry
    // points are named. Match by the DISTINCTIVE C function name (`sqlite3_*` / `git_*`) via the call's
    // LEAF — independent of the binding crate's alias: rusqlite calls `ffi::sqlite3_step`, git2 calls
    // `raw::git_remote_fetch`, and the nightly lint resolves the same to `libsqlite3_sys`/`libgit2_sys`;
    // all spellings share the leaf. Only the I/O-performing entry points are listed — the in-memory
    // accessors (`sqlite3_bind_*`/`sqlite3_column_*`, `git_*_oid`/strarray/options builders) stay pure,
    // so a non-listed `sqlite3_`/`git_` leaf returns None (under-report, never a wrong effect). Calibrated
    // + validated against rusqlite 0.39 / git2 0.20 source (eval/calibration).
    {
        let leaf = path.rsplit("::").next().unwrap_or(path);
        if let Some(rest) = leaf.strip_prefix("sqlite3_") {
            let _ = rest;
            // SQLite C API operations that touch the database (open/exec/step/prepare/backup/blob/wal).
            const DB: &[&str] = &[
                "sqlite3_open", "sqlite3_open_v2", "sqlite3_open16", "sqlite3_close", "sqlite3_close_v2",
                "sqlite3_exec", "sqlite3_step", "sqlite3_prepare", "sqlite3_prepare_v2",
                "sqlite3_prepare_v3", "sqlite3_prepare16", "sqlite3_prepare16_v2", "sqlite3_prepare16_v3",
                "sqlite3_get_table", "sqlite3_backup_init", "sqlite3_backup_step", "sqlite3_backup_finish",
                "sqlite3_blob_open", "sqlite3_blob_read", "sqlite3_blob_write", "sqlite3_blob_reopen",
                "sqlite3_load_extension", "sqlite3_wal_checkpoint", "sqlite3_wal_checkpoint_v2",
            ];
            return DB.contains(&leaf).then_some("Db");
        }
        if leaf.starts_with("git_") {
            // libgit2: remote/transport operations contact the network … (incl. submodule clone/update,
            // which `git_clone`/fetch the subrepo over its remote — `allow_fetch` defaults on; an A/B on
            // git2 0.20 caught `Submodule::update`/`clone` reporting no `Net`).
            const NET: &[&str] = &[
                "git_clone", "git_remote_connect", "git_remote_connect_ext", "git_remote_fetch",
                "git_remote_download", "git_remote_upload", "git_remote_push", "git_remote_ls",
                "git_submodule_clone", "git_submodule_update",
            ];
            // … and repository/index/odb/checkout/ref/config operations touch the on-disk .git store.
            const FS: &[&str] = &[
                "git_repository_open", "git_repository_open_ext", "git_repository_open_bare",
                "git_repository_init", "git_repository_init_ext", "git_repository_discover",
                "git_checkout_tree", "git_checkout_head", "git_checkout_index", "git_index_read",
                "git_index_write", "git_index_write_tree", "git_index_write_tree_to",
                "git_index_add_bypath", "git_index_add_all", "git_odb_open", "git_odb_read",
                "git_odb_write", "git_odb_open_wstream", "git_odb_open_rstream",
                "git_blob_create_fromdisk", "git_blob_create_fromworkdir", "git_blob_create_from_disk",
                "git_blob_create_from_workdir", "git_blob_create_from_stream", "git_commit_create",
                "git_commit_create_v", "git_reference_create", "git_reference_set_target",
                "git_reference_delete", "git_config_open_default", "git_config_open_ondisk",
                "git_config_add_file_ondisk", "git_tag_create", "git_treebuilder_write",
                "git_packbuilder_write",
            ];
            if NET.contains(&leaf) {
                return Some("Net");
            }
            if FS.contains(&leaf) {
                return Some("Fs");
            }
            return None;
        }
        if leaf.starts_with("curl_") {
            // libcurl (under the `curl` crate, called `curl_sys::curl_*`). Only the entry points that
            // PERFORM network I/O: the blocking transfer (`curl_easy_perform`), raw socket send/recv,
            // the HTTP/2 keepalive PING (`upkeep`), and the multi-interface transfer pumps. The large
            // pure surface (setopt/init/cleanup/reset/getinfo/escape/multi_add_handle/fdset/info_read)
            // stays unclassified, as do `curl_multi_wait`/`poll` (readiness WAIT on sockets, no payload —
            // the loop's `perform` is the tagged boundary, per the I/O-boundary principle). An A/B on
            // curl 0.4 caught the whole crate reporting ZERO Net (`Easy::perform` read as pure).
            const NET: &[&str] = &[
                "curl_easy_perform", "curl_easy_send", "curl_easy_recv", "curl_easy_upkeep",
                "curl_multi_perform", "curl_multi_socket_action",
            ];
            return NET.contains(&leaf).then_some("Net");
        }
        if let Some(op) = leaf.strip_prefix("SSL_") {
            // OpenSSL (libssl, under the `openssl`/`native-tls` crates, called `ffi::SSL_*`). The TLS
            // handshake and record I/O run over the peer socket -> Net. Unlike libc read/write, an SSL_*
            // op is ~always over a network BIO (the rare memory-BIO/sans-IO case is the honest exception
            // we accept). The crypto surface (EVP_*/SHA*/AES*) and pure setup (SSL_CTX_new/SSL_set_fd) are
            // NOT here; `BIO_*` is skipped (a BIO may be memory or socket). Validated vs openssl 0.9 source.
            const SSL_NET: &[&str] = &[
                "connect", "accept", "do_handshake", "read", "read_ex", "write", "write_ex", "peek",
                "peek_ex", "shutdown",
            ];
            return SSL_NET.contains(&op).then_some("Net");
        }
    }
    // HTTP clients use the same builder pattern as the AWS SDK: only the dispatch is
    // I/O. (Found by the eval: ebman's reqwest calls to the Anthropic API + webhooks
    // were silently classified network-free because reqwest wasn't recognized.)
    if crate_name == "reqwest" || crate_name == "isahc" {
        // The builder chain is pure; the dispatch (`::send`/`::execute`) is the I/O. PLUS the one-shot
        // CONVENIENCE functions `reqwest::get` / `reqwest::blocking::get` / `isahc::get`, which send
        // immediately — they're not the `Client::get` builder (a different path, `reqwest::Client::get`),
        // so an exact match avoids false-positiving the builder. (Found running on `xh`: a one-shot
        // `reqwest::get(url)` was classified network-free.)
        if path.ends_with("::send")
            || path.ends_with("::execute")
            || path == "reqwest::get"
            || path == "reqwest::blocking::get"
            || path == "isahc::get"
        {
            return Some("Net");
        }
        return None;
    }
    if crate_name == "ureq" && path.ends_with("::call") {
        return Some("Net");
    }
    // The `curl` crate (libcurl's safe binding — cargo's own HTTP client): the dispatch verbs are
    // `perform` (Easy/Easy2/Transfer/Multi), raw-socket `send`/`recv`, the keepalive `upkeep`, and the
    // multi-interface `action` (socket_action). The big setopt-style builder surface stays pure.
    // `Multi::timeout` is deliberately NOT matched: `Easy::timeout` is a pure CURLOPT_TIMEOUT setter
    // sharing the leaf — an under-report on the rare event-loop kick beats mis-tagging every consumer
    // that sets a timeout. (Consumer-side companion to the curl_* FFI tier, same A/B finding.)
    if crate_name == "curl"
        && (path.ends_with("::perform")
            || path.ends_with("::send")
            || path.ends_with("::recv")
            || path.ends_with("::upkeep")
            || path.ends_with("::action"))
    {
        return Some("Net");
    }
    // Message-queue clients fully encapsulate the socket (the underlying tokio::net lives
    // inside the crate, unseen), so a user's connect/publish/consume calls ARE the I/O
    // boundary — to a remote broker, hence Net. Match the broker round-trip verbs (snake_case
    // methods); the CamelCase option/property builders stay pure. (Found hardening on consumer
    // apps: lapin `basic_publish`/`queue_declare` and async-nats `publish`/`subscribe` were
    // classified pure — a message-queue client reporting no I/O.)
    if crate_name == "async_nats" {
        if path.ends_with("::connect")
            || path.contains("::publish")
            || path.ends_with("::subscribe")
            || path.ends_with("::queue_subscribe")
            || path.contains("::request")
            || path.ends_with("::flush")
        {
            return Some("Net");
        }
        return None;
    }
    if crate_name == "lapin" {
        if path.ends_with("::connect")
            || path.ends_with("::create_channel")
            || path.contains("::basic_")
            || path.contains("::queue_")
            || path.contains("::exchange_")
            || path.contains("::tx_")
            || path.ends_with("::confirm_select")
            || path.ends_with("::close")
        {
            return Some("Net");
        }
        return None;
    }
    // SMTP email — lettre's `Transport::send` is the network dispatch; Message building is
    // pure. (Found hardening on a lettre consumer: `mailer.send(&email)` classified pure.)
    if crate_name == "lettre" {
        if path.ends_with("::send") || path.ends_with("::send_raw") {
            return Some("Net");
        }
        return None;
    }
    // WebSockets — tungstenite (the modern successor to the old `websocket` crate). connect
    // and the socket read/write/send are network; Message constructors are pure. (Found on a
    // tungstenite consumer: connect + send + read classified pure.)
    if crate_name == "tungstenite" {
        if path.ends_with("::connect")
            || path.ends_with("::read")
            || path.ends_with("::write")
            || path.ends_with("::send")
            || path.ends_with("::close")
            || path.ends_with("::flush")
            || path.ends_with("::read_message")
            || path.ends_with("::write_message")
        {
            return Some("Net");
        }
        return None;
    }
    // elasticsearch: request builders are pure; only the `.send()` dispatch is HTTP I/O
    // (same shape as reqwest / the AWS SDK). (Found on an elasticsearch consumer.)
    if crate_name == "elasticsearch" && path.ends_with("::send") {
        return Some("Net");
    }
    // gRPC — tonic. The transport connect and the Grpc client RPC dispatch are network;
    // codecs and request/response wrappers are pure. (connect repro-confirmed on a consumer;
    // the unary/streaming RPC verbs are from the tonic::client::Grpc API.)
    if crate_name == "tonic" {
        if path.ends_with("::connect")
            || path.ends_with("::unary")
            || path.ends_with("::server_streaming")
            || path.ends_with("::client_streaming")
            || path.ends_with("::streaming")
        {
            return Some("Net");
        }
        return None;
    }
    // Kafka — rdkafka (FFI to librdkafka). Producer send + consumer poll/recv/subscribe/
    // commit are network round-trips to the brokers. (API-calibrated + unit-tested; a real
    // repro needs librdkafka/cmake, deferred.)
    if crate_name == "rdkafka" {
        if path.ends_with("::send")
            || path.ends_with("::send_result")
            || path.ends_with("::recv")
            || path.ends_with("::poll")
            || path.ends_with("::subscribe")
            || path.ends_with("::commit")
            || path.ends_with("::commit_message")
            || path.ends_with("::commit_consumer_state")
            || path.ends_with("::store_offset")
            || path.ends_with("::seek")
            || path.ends_with("::fetch_metadata")
            || path.ends_with("::fetch_watermarks")
            || path.ends_with("::flush")
        {
            return Some("Net");
        }
        return None;
    }
    // cap-std: capability-oriented std. I/O goes *through* a held capability handle
    // (Dir/Pool/Clock/...), so these calls ARE the effect. Recognising them means a
    // cap-std project's real I/O is detected and matches the capability it declared
    // (via `declared_caps`/`capstd_cap`) — conformance against unforgeable capabilities.
    if crate_name.starts_with("cap_") {
        if path.contains("::net::Unix") || path.contains("::os::") {
            return Some("Ipc");
        }
        if path.contains("::net") {
            return Some("Net");
        }
        if path.contains("::time") {
            return Some("Clock");
        }
        if path.contains("::fs") || crate_name == "cap_tempfile" || crate_name == "cap_directories" {
            return Some("Fs");
        }
        return None;
    }
    // Local IPC (Unix-domain sockets) is I/O but not *network* — keep it distinct so
    // CANDOR_NO_AMBIENT and audits don't conflate it with internet access. async-std puts its
    // Unix sockets under `os::unix::net` (mirroring std); async-net (smol's net layer) under
    // `unix`.
    if path.starts_with("tokio::net::Unix")
        || path.starts_with("std::os::unix::net")
        || path.starts_with("async_std::os::unix::net")
        || path.starts_with("async_net::unix")
    {
        return Some("Ipc");
    }
    // Raw packet capture / raw sockets — libpnet (the dominant low-level networking crate; powers
    // bandwhich, sniffers, custom-protocol tools). `datalink::channel` opens an L2 socket and
    // `transport::transport_channel` an L3/L4 raw socket — both ARE network I/O. Packet construction
    // (pnet_packet / pnet_base, MacAddr, Ethernet frames…) is pure and stays unclassified. The actual
    // frame read/write happens via methods on the returned Sender/Receiver (trait-object dispatch the
    // syntactic backend can't resolve), so the channel-open call is the precise Net boundary. (Found
    // scanning bandwhich — a packet sniffer — which reported Net 0.)
    if crate_name == "pnet" || crate_name == "pnet_datalink" || crate_name == "pnet_transport" {
        if path.ends_with("::channel") || path.ends_with("::transport_channel") {
            return Some("Net");
        }
        return None;
    }
    // Directory traversal — `ignore` (BurntSushi's gitignore-aware walker; powers ripgrep, fd). The walk
    // EXECUTORS read the directory tree from disk = Fs. Type-precise on purpose: the configuration builders
    // (`OverrideBuilder::build`, `GitignoreBuilder::build`, the `WalkBuilder` setters) and `DirEntry`
    // accessors are PURE — only `WalkBuilder::build`/`build_parallel` (which kick off the walk) and
    // `WalkParallel::run` (which drives it) touch the filesystem. A bare `build` would wrongly flag the
    // config builders. (Found scanning fd — a file finder — which reported Fs 2: its own `fs::read_dir`
    // was caught, but the `ignore`-based traversal that IS fd was invisible cross-crate.)
    if crate_name == "ignore" {
        if path == "ignore::WalkBuilder::build"
            || path == "ignore::WalkBuilder::build_parallel"
            || path.ends_with("::WalkParallel::run")
        {
            return Some("Fs");
        }
        return None;
    }
    // Filesystem watching — `notify` (the de-facto fs-watch crate: watchexec, cargo-watch, mdbook). A
    // watcher opens an OS notification handle (inotify / FSEvents / kqueue / ReadDirectoryChanges) and
    // registers paths — observing filesystem state changes = Fs. The lifecycle boundary: any
    // `*Watcher::new` constructor (RecommendedWatcher/PollWatcher/INotifyWatcher/FsEventWatcher/…), the
    // `recommended_watcher` convenience fn, and the `watch`/`unwatch` registration verbs. `Config`/`Event`/
    // `EventKind` data types stay pure. (Found scanning watchexec: its watcher-`create` read Fs 0.)
    if crate_name == "notify" {
        if path.ends_with("Watcher::new")
            || path.ends_with("::recommended_watcher")
            || path.ends_with("::watch")
            || path.ends_with("::unwatch")
        {
            return Some("Fs");
        }
        return None;
    }
    // Raw sockets. Match the I/O *types* only — `std::net` also holds pure data types
    // (SocketAddr, IpAddr, …) whose construction must NOT be flagged.
    if path.starts_with("std::net::TcpStream")
        || path.starts_with("std::net::TcpListener")
        || path.starts_with("std::net::UdpSocket")
        || path.starts_with("tokio::net::")
    {
        return Some("Net");
    }
    // Legacy tokio 0.1 socket crates — `tokio_tcp`/`tokio_udp` are *entirely* networking
    // (no pure types to over-flag), so the whole crate is Net. (Found hardening on websocat,
    // which is still on tokio 0.1: its `tokio_tcp::TcpStream::connect` was classified
    // network-free — a network tool confidently reporting 0 Net.)
    if matches!(crate_name, "tokio_tcp" | "tokio_udp") {
        return Some("Net");
    }
    // The other async runtimes mirror tokio's module layout, and their `net` modules hold only
    // socket I/O types (the pure `SocketAddr`/`IpAddr` are re-exports that resolve to `std::net`,
    // so they're excluded by def-path). `mio` is the low-level non-blocking-socket layer under
    // tokio/others; `async_net` is smol's net crate. Closes the async-std/smol/mio gap the
    // tokio_tcp note flagged. (Calibrated by module structure — these crates ARE networking — not
    // a live repro; the TCP/UDP types are defined in-crate so the def-path prefix is exact.)
    if path.starts_with("async_std::net::")
        || path.starts_with("mio::net::")
        || crate_name == "async_net"
    {
        return Some("Net");
    }
    // Database clients. Like the AWS/HTTP builders, only the execution verbs are I/O;
    // query *construction* is pure. Best-effort across crates (tune via CANDOR_CONFIG).
    // Note: bare `::query` is deliberately omitted — it executes in postgres/rusqlite but
    // only *builds* in sqlx, so including it would false-positive sqlx's `query()` builder.
    if DB_CRATES.contains(&crate_name) {
        // Postgres / SQLite-family clients: `query`/`batch_execute`/`prepare`/etc. ARE the
        // execution (round-trips to the server). sqlx is the outlier where bare `query()`
        // only BUILDS — it keeps the narrow set below. (Found by running on a real
        // tokio-postgres app, pgman: candor had reported only 4 of ~20 DB call sites.)
        if matches!(crate_name, "postgres" | "tokio_postgres" | "deadpool_postgres" | "rusqlite") {
            const PG: [&str; 19] = [
                "::query", "::query_one", "::query_opt", "::query_raw", "::execute",
                "::batch_execute", "::simple_query", "::prepare", "::prepare_typed",
                "::copy_in", "::copy_out", "::transaction", "::connect",
                // rusqlite's dialect of the same verbs (a verb-probe found the CANONICAL rusqlite
                // consumer API classifying pure): `query_row` is the one-row read, `query_map`/
                // `query_and_then` the many-row reads, `execute_batch` is rusqlite's name for
                // batch_execute, `prepare_cached` round-trips like prepare. `query_typed` is
                // tokio_postgres 0.7.10+.
                "::query_row", "::query_map", "::query_and_then", "::execute_batch",
                "::prepare_cached", "::query_typed",
            ];
            if PG.iter().any(|v| path.ends_with(v)) {
                return Some("Db");
            }
            // rusqlite only: opening the database IS the connection establishment (`Connection::
            // open`/`open_in_memory`/`open_with_flags` — the embedded analog of `::connect`).
            if crate_name == "rusqlite"
                && (path.ends_with("::open")
                    || path.ends_with("::open_in_memory")
                    || path.ends_with("::open_with_flags"))
            {
                return Some("Db");
            }
            return None;
        }
        // redis: the way redis is ACTUALLY used is the high-level `Commands`/`AsyncCommands`
        // traits (`con.get`/`set`/`hset`/`lpush`/…) — every method is a round-trip — plus
        // connection establishment. The shared VERBS below only catch the low-level
        // `cmd("GET").query(con)`, so without this a normal redis user's calls classify as
        // PURE. (Found hardening on redis-rs: a fn doing `con.get`/`set` reported no effects.)
        if crate_name == "redis"
            && (path.contains("Commands::")
                || path.contains("::get_connection")
                || path.contains("::get_async_connection")
                || path.contains("::get_multiplexed_async_connection")
                || path.contains("ConnectionManager")
                || path.ends_with("::query")
                || path.ends_with("::query_async")
                || path.ends_with("::req_command")
                || path.ends_with("::req_packed_command")
                || path.ends_with("::req_packed_commands"))
        {
            return Some("Db");
        }
        // mongodb: a document-store API with none of the SQL verbs — the user calls
        // `coll.find_one`/`insert_one`/`aggregate`/… and `Client::with_uri_str`. Without
        // these a mongodb user's calls classify PURE. (Found hardening: a fn doing
        // `find_one`+`insert_one` reported no effects.) Handle accessors (name/namespace)
        // and option/doc builders don't match these verbs, so they stay pure.
        if crate_name == "mongodb" {
            const MONGO: [&str; 27] = [
                "::with_uri_str", "::connect", "::find", "::find_one", "::insert_one",
                "::insert_many", "::update_one", "::update_many", "::delete_one",
                "::delete_many", "::replace_one", "::aggregate", "::count_documents",
                "::estimated_document_count", "::count", "::distinct", "::run_command",
                "::find_one_and_update", "::find_one_and_delete", "::find_one_and_replace",
                "::list_collections", "::list_collection_names", "::list_databases",
                "::list_database_names", "::create_collection", "::create_index", "::watch",
            ];
            if MONGO.iter().any(|v| path.ends_with(v)) {
                return Some("Db");
            }
            return None;
        }
        // mysql / mysql_async: the `query`/`exec` families + `get_conn`/`ping` execute
        // immediately — no build-then-execute split like sqlx, so matching `::query` is safe
        // here. Same DB-verb-dialect gap class as redis/mongodb; calibrated from the Queryable
        // API (unit-tested; a real-app repro is the remaining confirmation).
        if matches!(crate_name, "mysql" | "mysql_async") {
            const MY: [&str; 16] = [
                "::query", "::query_first", "::query_iter", "::query_map", "::query_fold",
                "::query_drop", "::exec", "::exec_first", "::exec_iter", "::exec_map",
                "::exec_fold", "::exec_drop", "::exec_batch", "::prep", "::ping", "::get_conn",
            ];
            if MY.iter().any(|v| path.ends_with(v)) {
                return Some("Db");
            }
            return None;
        }
        // sea_orm: an ORM whose execution is split from building (like sqlx). The query
        // BUILDERS (`Entity::find`, `Entity::insert`) are pure; execution happens at `.all`/
        // `.one`/`.count`/`.stream` and `Insert/Update/Delete::exec`. The write path via an
        // ActiveModel (`model.insert(db)`) executes too — distinguished from the `EntityTrait`
        // builder by the trait in the path (`ActiveModelTrait::`). (Found hardening on a
        // sea_orm consumer app: `.all(db)` reads and `ActiveModel::insert` writes were pure.)
        if crate_name == "sea_orm" {
            if path.ends_with("::all")
                || path.ends_with("::one")
                || path.ends_with("::count")
                || path.ends_with("::stream")
                || path.ends_with("::exec")
                || path.ends_with("::exec_with_returning")
                || path.ends_with("::exec_without_returning")
                || path.ends_with("::connect")
                || path.ends_with("::execute")
                || path.ends_with("::execute_unprepared")
                || path.ends_with("::query_one")
                || path.ends_with("::query_all")
                || path.ends_with("::fetch_page")
                || path.ends_with("::num_items")
                || path.contains("ActiveModelTrait::")
            {
                return Some("Db");
            }
            return None;
        }
        // (Reached by sqlx + diesel — the build-vs-execute-split crates.) `first` is diesel's
        // LIMIT-1 round trip and `load_iter` its 2.x streaming execution; `fetch_many` is sqlx's
        // multi-result stream. All crate-gated, so a std `Vec::first` never resolves here.
        const VERBS: [&str; 19] = [
            "::execute", "::query_row", "::query_map", "::query_one", "::fetch_one",
            "::fetch_all", "::fetch_optional", "::fetch", "::fetch_many", "::connect",
            "::acquire", "::begin", "::commit", "::rollback", "::load", "::load_iter",
            "::first", "::get_result", "::get_results",
        ];
        if VERBS.iter().any(|v| path.ends_with(v)) {
            return Some("Db");
        }
        return None;
    }
    // std::path::Path / PathBuf STAT-family methods hit the filesystem (each is a stat/readlink/
    // readdir syscall) — unlike the rest of the std::path surface, which is pure string manipulation
    // (join/file_name/extension/parent/…). Verb-precise so the scanner's receiver inference can safely
    // route a `path.symlink_metadata()` method call here. (A blackout screen caught gix-dir — an entire
    // directory WALKER — reporting ZERO Fs because all its I/O is Path-method calls; same class as
    // fd's residual `Path::symlink_metadata` under-report.)
    if let Some(m) = path
        .strip_prefix("std::path::Path::")
        .or_else(|| path.strip_prefix("std::path::PathBuf::"))
    {
        const STAT: &[&str] = &[
            "metadata", "symlink_metadata", "canonicalize", "read_link", "read_dir", "exists",
            "try_exists", "is_file", "is_dir", "is_symlink",
        ];
        return STAT.contains(&m).then_some("Fs");
    }
    // Filesystem. `tokio::fs`/`async_std::fs` are the async mirrors of `std::fs`; `async_fs` is
    // smol's fs crate; `fs_err` is a drop-in `std::fs` wrapper (its whole surface is fs I/O).
    if path.starts_with("std::fs::")
        || path.starts_with("tokio::fs::")
        || path.starts_with("async_std::fs::")
        || crate_name == "async_fs"
        || crate_name == "fs_err"
        || crate_name == "memmap2"
    {
        return Some("Fs");
    }
    // tempfile: creating a temp file/dir touches the disk. Match the create/persist verbs (the
    // `Builder` setters — prefix/suffix/rand_bytes — stay pure). `persist`/`keep` rename/retain
    // the file on disk; `close` removes it.
    if crate_name == "tempfile"
        && (path.ends_with("::tempfile")
            || path.ends_with("::tempfile_in")
            || path.ends_with("::tempdir")
            || path.ends_with("::tempdir_in")
            || path.ends_with("NamedTempFile::new")
            || path.ends_with("NamedTempFile::new_in")
            || path.ends_with("TempDir::new")
            || path.ends_with("TempDir::new_in")
            || path.ends_with("::persist")
            || path.ends_with("::persist_noclobber")
            || path.ends_with("::keep"))
    {
        return Some("Fs");
    }
    // glob: walks the filesystem to expand a pattern (the returned iterator reads directories).
    // `Pattern::matches` is pure string matching — match only the directory-walking entry points.
    if crate_name == "glob" && (path.ends_with("::glob") || path.ends_with("::glob_with")) {
        return Some("Fs");
    }
    // Randomness / entropy. `getrandom`/`fastrand` are effectful end-to-end. `rand` is NOT — it
    // mixes entropy/generation (effectful) with *pure* distribution constructors (`Uniform::new`,
    // `Normal::new`) and deterministic-seed constructors (`seed_from_u64`). Flagging the whole crate
    // over-reported those as `Rand`; match only the calls that actually consume randomness — the
    // entropy sources (`OsRng`, `thread_rng`/`rng`, `from_entropy`/`from_os_rng`) and the generation
    // verbs (`gen*`/`random*`/`fill*`/`sample*`/`next_u*`). A `Uniform::new` is now correctly pure.
    if crate_name == "getrandom" || crate_name == "fastrand" {
        return Some("Rand");
    }
    if crate_name == "rand" {
        let rng_verb = path.ends_with("::gen")
            || path.ends_with("::gen_range")
            || path.ends_with("::gen_bool")
            || path.ends_with("::gen_ratio")
            || path.ends_with("::random")
            || path.ends_with("::random_range")
            || path.ends_with("::random_bool")
            || path.ends_with("::random_ratio")
            || path.ends_with("::random_iter") // rand 0.9 iterator generator
            || path.ends_with("::gen_iter")
            || path.ends_with("::fill")
            || path.ends_with("::fill_bytes")
            || path.ends_with("::try_fill")
            || path.ends_with("::try_fill_bytes")
            || path.ends_with("::sample")
            || path.ends_with("::sample_iter")
            || path.ends_with("::next_u32")
            || path.ends_with("::next_u64")
            || path.ends_with("::thread_rng")
            || path.ends_with("::rng")
            || path.ends_with("::from_entropy")
            || path.ends_with("::from_os_rng");
        if rng_verb || path.contains("OsRng") {
            return Some("Rand");
        }
        return None;
    }
    // Subprocess spawning. `tokio::process` is the async mirror of `std::process` — it exists
    // only to spawn/control subprocesses (`Command`/`Child`, no pure data types like std's
    // `Stdio`/`ExitStatus`/`exit`), so spawning through it is Exec just the same. Without this an
    // async app's `tokio::process::Command::new(..).spawn()` classified pure — a silent under-report
    // of subprocess execution, the dangerous direction (mirrors the tokio::fs/tokio::net coverage).
    if path.starts_with("std::process::Command")
        || path.starts_with("std::process::Child")
        || path.starts_with("tokio::process::Command")
        || path.starts_with("tokio::process::Child")
        || path.starts_with("async_std::process::Command")
        || path.starts_with("async_std::process::Child")
        || crate_name == "async_process"
        || crate_name == "portable_pty"
    {
        return Some("Exec");
    }
    // duct: a subprocess-orchestration crate. `cmd()`/`cmd!` only *build* an Expression; the
    // spawn/wait happens at `run`/`read`/`start`. Match the execution verbs, not the builder.
    if crate_name == "duct"
        && (path.ends_with("::run")
            || path.ends_with("::read")
            || path.ends_with("::start")
            || path.ends_with("::read_chars"))
    {
        return Some("Exec");
    }
    if path.starts_with("std::env::") {
        return Some("Env");
    }
    // dotenvy / dotenv: load environment variables (reading a `.env` file and mutating the process
    // environment). Match the load/read entry points; `Error`/builder types stay pure.
    if matches!(crate_name, "dotenvy" | "dotenv")
        && (path.ends_with("::dotenv")
            || path.ends_with("::dotenv_override")
            || path.ends_with("::from_path")
            || path.ends_with("::from_path_override")
            || path.ends_with("::from_filename")
            || path.ends_with("::from_filename_override")
            || path.ends_with("::from_read")
            || path.ends_with("::from_read_override")
            || path.ends_with("::load")
            || path.ends_with("::var")
            || path.ends_with("::vars"))
    {
        return Some("Env");
    }
    // Wall-clock reads. Match the `now` accessor precisely (ends_with), not any path
    // containing the substring "now". The `time` crate (distinct from `std::time`/`chrono`)
    // reads the clock via `now_utc`/`now_local` (and the deprecated `Instant::now`).
    if (crate_name == "chrono" || path.starts_with("std::time::")) && path.ends_with("::now") {
        return Some("Clock");
    }
    if crate_name == "time"
        && (path.ends_with("::now_utc") || path.ends_with("::now_local") || path.ends_with("::now"))
    {
        return Some("Clock");
    }
    if crate_name == "tracing" {
        return Some("Log");
    }
    // The `log` facade: its macros route through `log::__private_api`; the crate's types
    // (`Level`, `LevelFilter`) are pure, so match the logging entry, not the whole crate.
    if crate_name == "log" && path.contains("::__private_api") {
        return Some("Log");
    }
    // Compiler diagnostic emission — the ONE genuinely effectful operation in the otherwise-pure
    // rustc_* surface (a dylint lint's actual OUTPUT: it writes warnings/errors to the compiler's
    // diagnostic sink). Classified `Log` (same family as `tracing`/`log` — program output). Match the
    // emission verbs precisely; rustc_lint/rustc_errors are mostly pure types (Lint, LintId, the Diag
    // BUILDERS), and only the terminal `emit`/`emit_span_lint` actually produces output.
    if crate_name == "rustc_lint"
        && (path.ends_with("::emit_span_lint")
            || path.ends_with("::span_lint")
            || path.ends_with("::span_lint_hir"))
    {
        return Some("Log");
    }
    if crate_name == "rustc_errors"
        && (path.ends_with("::emit")
            || path.ends_with("::emit_diagnostic")
            || path.ends_with("::emit_now"))
    {
        return Some("Log");
    }
    if crate_name == "arboard" {
        return Some("Clipboard");
    }
    None
}

pub fn cap_from_name(name: &str) -> Option<&'static str> {
    EFFECTS.iter().copied().find(|e| *e == name)
}

/// Map a cap-std capability *type* to the effect it authorises. Holding one of these
/// (e.g. `&Dir`) is the real, unforgeable right to perform that effect — so candor
/// treats it as a declared capability, exactly like its own `&Fs` token.
pub fn capstd_cap(crate_name: &str, type_name: &str) -> Option<&'static str> {
    if !crate_name.starts_with("cap_") {
        return None;
    }
    Some(match type_name {
        "Dir" => "Fs",
        "TcpListener" | "TcpStream" | "UdpSocket" | "Pool" => "Net",
        "UnixListener" | "UnixStream" | "UnixDatagram" => "Ipc",
        "SystemClock" | "MonotonicClock" => "Clock",
        _ => return None,
    })
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn db_crates_are_calibrated() {
        // The calibrated set must cover every DB client the classifier knows, or the receipt's coverage
        // check would flag a recognized crate as a blind spot. (Was nightly-lint-only; now runs on stable.)
        for c in DB_CRATES {
            assert!(
                CALIBRATED_CRATES.contains(&c),
                "DB crate `{c}` is matched by classify() but missing from CALIBRATED_CRATES"
            );
        }
    }

    #[test]
    fn calibrated_crates_are_live() {
        // Conversely, every crate advertised as calibrated must actually be matched by classify() for
        // some representative path — a dead entry would silently suppress a real coverage warning.
        for c in CALIBRATED_CRATES {
            assert!(
                CALIBRATION_PROBE_TAILS.iter().any(|t| classify(c, &format!("{c}{t}")).is_some()),
                "calibrated crate `{c}` is matched by no path in classify() — dead list entry"
            );
        }
    }

    #[test]
    fn classify_core_effects() {
        // A representative smoke test of the classifier's main families, so the published crate is not
        // shipped untested (these used to live only in the nightly-only src/lib.rs).
        assert_eq!(classify("std", "std::fs::read_to_string"), Some("Fs"));
        // std::path stat-family methods are Fs (each is a stat/readdir syscall); the pure
        // string-manipulation surface stays unclassified (the blackout screen's gix-dir find).
        assert_eq!(classify("std", "std::path::Path::symlink_metadata"), Some("Fs"));
        assert_eq!(classify("std", "std::path::PathBuf::read_dir"), Some("Fs"));
        assert_eq!(classify("std", "std::path::Path::exists"), Some("Fs"));
        assert_eq!(classify("std", "std::path::Path::join"), None); // pure string manipulation
        assert_eq!(classify("std", "std::path::PathBuf::file_name"), None);
        assert_eq!(classify("std", "std::path::Path::parent"), None);
        assert_eq!(classify("std", "std::process::Command::new"), Some("Exec"));
        assert_eq!(classify("std", "std::env::var"), Some("Env"));
        assert_eq!(classify("reqwest", "reqwest::Client::execute"), Some("Net"));
        // one-shot convenience fns send immediately → Net; the `Client::get` builder stays pure.
        assert_eq!(classify("reqwest", "reqwest::get"), Some("Net"));
        assert_eq!(classify("reqwest", "reqwest::blocking::get"), Some("Net"));
        assert_eq!(classify("reqwest", "reqwest::Client::get"), None);
        assert_eq!(classify("reqwest", "reqwest::RequestBuilder::header"), None);
        // nix routes through the libc syscall table (same leaves): I/O classified, generic fd ops skipped.
        assert_eq!(classify("nix", "nix::fcntl::open"), Some("Fs"));
        assert_eq!(classify("nix", "nix::sys::socket::connect"), Some("Net"));
        assert_eq!(classify("nix", "nix::unistd::execvp"), Some("Exec"));
        assert_eq!(classify("nix", "nix::unistd::write"), None); // generic fd op — deliberately unclassified
        assert_eq!(classify("nix", "nix::unistd::getpid"), None); // not I/O
        // rustix does raw syscalls (no libc underneath) → classified directly by leaf, same table.
        assert_eq!(classify("rustix", "rustix::time::clock_settime"), Some("Clock"));
        assert_eq!(classify("rustix", "rustix::fs::symlink"), Some("Fs"));
        assert_eq!(classify("rustix", "rustix::net::connect"), Some("Net"));
        assert_eq!(classify("rustix", "rustix::io::read"), None); // generic fd op
        // pnet raw packet capture: channel openers are Net, packet construction stays pure.
        assert_eq!(classify("pnet", "pnet::datalink::channel"), Some("Net"));
        assert_eq!(classify("pnet", "pnet::transport::transport_channel"), Some("Net"));
        assert_eq!(classify("pnet_datalink", "pnet_datalink::channel"), Some("Net"));
        assert_eq!(classify("pnet", "pnet::packet::ethernet::EthernetPacket::new"), None);
        assert_eq!(classify("pnet_base", "pnet_base::MacAddr::new"), None);
        // ignore (gitignore-aware walker): walk executors are Fs, config builders stay pure.
        assert_eq!(classify("ignore", "ignore::WalkBuilder::build_parallel"), Some("Fs"));
        assert_eq!(classify("ignore", "ignore::WalkBuilder::build"), Some("Fs"));
        assert_eq!(classify("ignore", "ignore::WalkParallel::run"), Some("Fs"));
        assert_eq!(classify("ignore", "ignore::overrides::OverrideBuilder::build"), None); // pure config
        assert_eq!(classify("ignore", "ignore::gitignore::GitignoreBuilder::build"), None); // pure config
        assert_eq!(classify("ignore", "ignore::DirEntry::path"), None); // pure accessor
        // notify fs-watching: watcher constructors + watch/unwatch are Fs, data types stay pure.
        assert_eq!(classify("notify", "notify::RecommendedWatcher::new"), Some("Fs"));
        assert_eq!(classify("notify", "notify::PollWatcher::new"), Some("Fs"));
        assert_eq!(classify("notify", "notify::recommended_watcher"), Some("Fs"));
        assert_eq!(classify("notify", "notify::INotifyWatcher::watch"), Some("Fs"));
        assert_eq!(classify("notify", "notify::Config::default"), None); // pure config
        assert_eq!(classify("notify", "notify::Event::new"), None); // pure data type
        assert_eq!(classify("rusqlite", "rusqlite::Connection::execute"), Some("Db"));
        // the rusqlite verb DIALECT (a verb probe found the canonical consumer API classifying pure):
        assert_eq!(classify("rusqlite", "rusqlite::Connection::query_row"), Some("Db"));
        assert_eq!(classify("rusqlite", "rusqlite::Statement::query_map"), Some("Db"));
        assert_eq!(classify("rusqlite", "rusqlite::Connection::execute_batch"), Some("Db"));
        assert_eq!(classify("rusqlite", "rusqlite::Connection::prepare_cached"), Some("Db"));
        assert_eq!(classify("rusqlite", "rusqlite::Connection::open"), Some("Db"));
        assert_eq!(classify("rusqlite", "rusqlite::Connection::open_in_memory"), Some("Db"));
        // …but `open` stays rusqlite-only (postgres has no open; nothing else may borrow it):
        assert_eq!(classify("postgres", "postgres::Client::open"), None);
        assert_eq!(classify("tokio_postgres", "tokio_postgres::Client::query_typed"), Some("Db"));
        // diesel's LIMIT-1 + streaming executions; sqlx's multi-result stream:
        assert_eq!(classify("diesel", "diesel::RunQueryDsl::first"), Some("Db"));
        assert_eq!(classify("diesel", "diesel::RunQueryDsl::load_iter"), Some("Db"));
        assert_eq!(classify("sqlx", "sqlx::query::Query::fetch_many"), Some("Db"));
        // sqlx's bare `query()` builder must STAY pure (the original sqlx lesson):
        assert_eq!(classify("sqlx", "sqlx::query"), None);
        assert_eq!(classify("tracing", "tracing::event"), Some("Log"));
        // FFI tiers (matched by distinctive leaf, alias-independent)
        assert_eq!(classify("libc", "libc::open"), Some("Fs"));
        assert_eq!(classify("libc", "libc::connect"), Some("Net"));
        assert_eq!(classify("libc", "libc::read"), None); // generic fd op — deliberately unclassified
        assert_eq!(classify("ffi", "ffi::sqlite3_step"), Some("Db"));
        assert_eq!(classify("raw", "raw::git_remote_fetch"), Some("Net"));
        // libgit2 clone + submodule clone/update fetch over the network (an A/B on git2 0.20 caught
        // `Submodule::update`/`clone` and `Repository::clone` reporting no Net — the latter because the
        // `src/build.rs` module was being dropped as if it were the Cargo build script).
        assert_eq!(classify("raw", "raw::git_clone"), Some("Net"));
        assert_eq!(classify("raw", "raw::git_submodule_clone"), Some("Net"));
        assert_eq!(classify("raw", "raw::git_submodule_update"), Some("Net"));
        assert_eq!(classify("raw", "raw::git_submodule_open"), None); // local subrepo open — not Net
        // libcurl: the transfer/raw-socket entry points are Net (an A/B on curl 0.4 caught the whole
        // crate reporting ZERO Net); the big setopt/init/getinfo surface — and the readiness-wait
        // multi_wait/poll — stay unclassified (the loop's perform is the boundary).
        assert_eq!(classify("curl_sys", "curl_sys::curl_easy_perform"), Some("Net"));
        assert_eq!(classify("curl_sys", "curl_sys::curl_easy_send"), Some("Net"));
        assert_eq!(classify("curl_sys", "curl_sys::curl_multi_perform"), Some("Net"));
        assert_eq!(classify("curl_sys", "curl_sys::curl_multi_socket_action"), Some("Net"));
        assert_eq!(classify("curl_sys", "curl_sys::curl_easy_setopt"), None); // in-memory option write
        assert_eq!(classify("curl_sys", "curl_sys::curl_easy_init"), None); // handle alloc
        assert_eq!(classify("curl_sys", "curl_sys::curl_multi_wait"), None); // readiness wait, no payload
        // consumer-side `curl` crate rule: the dispatch verbs are Net, the setopt builders pure.
        assert_eq!(classify("curl", "curl::easy::Easy::perform"), Some("Net"));
        assert_eq!(classify("curl", "curl::multi::Multi::perform"), Some("Net"));
        assert_eq!(classify("curl", "curl::easy::Easy::send"), Some("Net"));
        assert_eq!(classify("curl", "curl::easy::Easy::url"), None); // CURLOPT setter — pure
        assert_eq!(classify("curl", "curl::easy::Easy::timeout"), None); // pure setter; Multi::timeout under-reported by design
        assert_eq!(classify("ffi", "ffi::SSL_connect"), Some("Net"));
        // pure crates stay pure
        assert_eq!(classify("serde", "serde::Serialize::serialize"), None);
        assert_eq!(classify("std", "std::vec::Vec::push"), None);
    }
}