Module diagnostics 

Per-take diagnostics written to disk, best-effort.

One take becomes one directory holding the original capture, every transcription stage, and the timings - so “which stage ate the words” is answerable by opening files rather than by rerunning the take under investigation. The layout is granular on purpose: the pre-normalization text is kept per window, because the joined form cannot show whether a window lost the words or the join did.

Everything here is best-effort: a capture that cannot be written is logged and dropped, never propagated, because diagnostics must never break a take.

Retention: the store keeps the last RETAINED_TAKES takes, pruning the oldest by directory name after each write.

Re-exports

use std::path::Path; 🔒

use std::path::PathBuf; 🔒

use std::time::SystemTime; 🔒

use std::time::UNIX_EPOCH; 🔒

use serde_json::json; 🔒

use transcribe_cpp::CancelToken; 🔒

use crate::audio::SAMPLE_RATE; 🔒

use crate::config::ModelSpec; 🔒

use crate::engine::Stages; 🔒

Structs

TakeModels 🔒

The models a take ran under: the resolved specs the engine was built with.

TakeRecord 🔒

Everything one finished take contributes to the store.

WindowRecord 🔒

One window’s record: the slice of capture it covered and the raw pre-normalization text it produced.

Constants

RETAINED_TAKES 🔒

How many takes the store keeps, and the bench’s corpus along with it.

Functions

capture_take 🔒

Write one take into dir/take-<take_id>/ and prune the store to RETAINED_TAKES. Nothing here can fail the caller: every step logs its own failure and stops that take’s capture.

claim_take_dir 🔒

Claim dir/take-<stamp>/, advancing a millisecond per name already taken, and answer the name and the directory it claimed.

model_identity 🔒

One role’s resolved spec as meta.json names it: the file, the digest it declares where it declares one, and the size.

prune 🔒

Delete every take but the newest RETAINED_TAKES. Only store directories count - take- followed by nothing but digits - so a user directory that happens to share the prefix is never touched. Removal is best-effort, and a directory that will not leave is logged: it occupies a retention slot invisibly otherwise.

sha256_file 🔒

A file’s sha256, lowercase hex; empty when it cannot be read.

starts_take 🔒

A store directory: take- followed by all-ASCII digits. Anything else - take-2026-photos, a user’s own folder - is not ours to prune.

take_stamp 🔒

The unix-millisecond stamp a take directory is named by, as take-<13 digits>: sortable, and 13 digits holds until the year 2286 so lexicographic order never lies about recency.

write_take 🔒

One claimed take’s files, meta.json last, answering whether the take was written whole.

write_wav 🔒

Encode audio as a canonical 16-bit PCM wav at the one rate every model here reads.