fix(backfill): convert exiftool-legacy rows whose only dates are zero

backfill_utc_dates refused any fs_time re-resolution of a tag-dated row,
assuming exiftool had failed on the file. Files whose date tags are all
0000:00:00 (e.g. uv-prod-long.mp4) were read fine and simply have no
date, so they sat unconverted and re-ran through exiftool every tick.

resolve_dates_batch_report now also returns the paths exiftool read
without error (-Error surfaces ExifTool:Error in the JSON). The drain
lets fs_time stand for an exiftool-legacy row only when exiftool read
the file; read errors / missing entries and kamadak exif-legacy rows
keep refusing.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

CLAUDE.md

@@ -458,7 +458,10 @@ the file. Missing-file rows are paged past without spending the cap, and
 a per-library resume cursor (carried across ticks by the watcher, wraps
 at end of table) keeps unconvertible rows from stalling the rows behind
 them; EXIF/exiftool-legacy rows wait when exiftool is absent (they'd
-otherwise fall to fs_time). When a converted row's display-zone day differs from
+otherwise fall to fs_time), and a re-resolution to fs_time never replaces
+a tag date — except an `exiftool`-legacy row that exiftool read cleanly
+and found no usable date in (all-zero tags), where fs_time is the honest
+answer (`resolve_dates_batch_report`'s `exiftool_read`). When a converted row's display-zone day differs from
 the day its legacy value rendered on, the `day`-span `precomputed_reels`
 rows covering either month-day (this library + `all`) are deleted and
 unlinked; the nightly pre-gen regenerates. Daily conversation summaries

src/backfill.rs

@@ -302,6 +302,32 @@ pub fn backfill_utc_dates(
     backfill_utc_dates_capped(library, exif_dao, reel_dao, display_tz, cursor, cap)
 }
 
+/// Whether a re-resolution can stand in for a row's legacy date. Only an
+/// `fs_time` landing on a tag-dated row is suspect:
+///
+/// - legacy `exif` (kamadak-exif) — a different reader found a date there,
+///   so exiftool coming up empty says nothing about it: refuse.
+/// - legacy `exiftool` — refuse unless exiftool read the file this pass.
+///   If it did and found no usable date (e.g. a muxer's all-zero
+///   `CreateDate`s), the legacy value was never a real capture date and
+///   fs_time is the honest answer; without this the row re-runs through
+///   exiftool every tick and never converts. If it didn't (crash, read
+///   error), the legacy date may be real.
+fn utc_resolution_is_faithful(
+    legacy_source: Option<&str>,
+    resolved_source: date_resolver::DateSource,
+    exiftool_read: bool,
+) -> bool {
+    if resolved_source != date_resolver::DateSource::FsTime {
+        return true;
+    }
+    match legacy_source {
+        Some("exif") => false,
+        Some("exiftool") => exiftool_read,
+        _ => true,
+    }
+}
+
 fn backfill_utc_dates_capped(
     library: &libraries::Library,
     exif_dao: &Arc<Mutex<Box<dyn ExifDao>>>,
@@ -390,14 +416,16 @@ fn backfill_utc_dates_capped(
         // One exiftool subprocess for the whole batch — byte-identical
         // call shape to the #117 audit preview.
         let paths: Vec<PathBuf> = filed.iter().map(|(_, p)| p.clone()).collect();
-        let resolved = date_resolver::resolve_dates_batch(&paths, &HashMap::new());
+        let batch = date_resolver::resolve_dates_batch_report(&paths, &HashMap::new());
         for (row, abs) in filed {
-            let legacy_had_tag = matches!(
-                row.date_taken_source.as_deref(),
-                Some("exif") | Some("exiftool")
-            );
-            match resolved.get(&abs) {
-                Some(rd) if !(legacy_had_tag && rd.source == date_resolver::DateSource::FsTime) => {
+            match batch.resolved.get(&abs) {
+                Some(rd)
+                    if utc_resolution_is_faithful(
+                        row.date_taken_source.as_deref(),
+                        rd.source,
+                        batch.exiftool_read.contains(&abs),
+                    ) =>
+                {
                     match (rd.date_utc, rd.tz_source) {
                         (Some(utc), Some(tz)) => conversions.push(Conversion {
                             row,
@@ -408,9 +436,9 @@ fn backfill_utc_dates_capped(
                         _ => outcome.unresolved += 1,
                     }
                 }
-                // No result, or a row whose legacy date came from a tag
-                // now resolving to fs_time (exiftool failed on it): don't
-                // bake a filesystem time in over a real capture date.
+                // No result, or a tag-dated row now resolving to fs_time
+                // without exiftool having read the file: don't bake a
+                // filesystem time in over a real capture date.
                 _ => {
                     debug!(
                         "utc_backfill: no faithful resolution for lib {} {}",
@@ -1494,6 +1522,111 @@ mod tests {
         assert_eq!(progress[&1], 0);
     }
 
+    #[test]
+    fn utc_resolution_fs_time_over_tag_date_needs_an_exiftool_read() {
+        use date_resolver::DateSource::{Exiftool, FsTime};
+        // Anything but fs_time stands, whatever the legacy source.
+        assert!(utc_resolution_is_faithful(
+            Some("exiftool"),
+            Exiftool,
+            false
+        ));
+        assert!(utc_resolution_is_faithful(Some("exif"), Exiftool, false));
+        // fs_time over a non-tag legacy source stands.
+        assert!(utc_resolution_is_faithful(Some("fs_time"), FsTime, false));
+        assert!(utc_resolution_is_faithful(None, FsTime, false));
+        // Legacy exiftool: only once exiftool read the file and found no
+        // usable date.
+        assert!(utc_resolution_is_faithful(Some("exiftool"), FsTime, true));
+        assert!(!utc_resolution_is_faithful(Some("exiftool"), FsTime, false));
+        // Legacy kamadak exif: never — exiftool's read doesn't vouch for it.
+        assert!(!utc_resolution_is_faithful(Some("exif"), FsTime, true));
+        assert!(!utc_resolution_is_faithful(Some("exif"), FsTime, false));
+    }
+
+    /// Tiny H.264 MP4 with no metadata: exiftool reads it cleanly but every
+    /// QuickTime / Track date is the `0000:00:00 00:00:00` placeholder.
+    fn write_zero_date_video(root: &std::path::Path, rel: &str) {
+        run_ok(
+            std::process::Command::new("ffmpeg")
+                .args([
+                    "-y",
+                    "-loglevel",
+                    "error",
+                    "-f",
+                    "lavfi",
+                    "-i",
+                    "color=c=black:s=32x32:d=0.4",
+                    "-c:v",
+                    "libx264",
+                    "-pix_fmt",
+                    "yuv420p",
+                    "-fflags",
+                    "+bitexact",
+                    "-map_metadata",
+                    "-1",
+                ])
+                .arg(root.join(rel)),
+        );
+    }
+
+    #[test]
+    fn utc_backfill_legacy_exiftool_row_with_only_zero_dates_converts_to_fs_time() {
+        if !tools_available() {
+            eprintln!("skipping: ffmpeg/exiftool not on PATH");
+            return;
+        }
+        let (tmp, library, conn, exif_dao, _face_dao) = setup();
+        let reel_dao = reel_dao_for(&conn);
+        // Both legacy-exiftool rows land on fs_time. exiftool reads the
+        // video cleanly and finds only zero dates (the prod
+        // `uv-prod-long.mp4` case); the "jpg" is binary junk exiftool
+        // reports a File format error for, so it never got a look inside.
+        // (Printable junk wouldn't do: exiftool reads it as a text file.)
+        write_zero_date_video(tmp.path(), "uv-long.mp4");
+        write_image(
+            tmp.path(),
+            "unreadable.jpg",
+            &[0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
+        );
+        let video = insert_legacy(
+            &exif_dao,
+            1,
+            "uv-long.mp4",
+            Some(1_000_000_000),
+            Some("exiftool"),
+        );
+        let junk = insert_legacy(
+            &exif_dao,
+            1,
+            "unreadable.jpg",
+            Some(1_000_000_000),
+            Some("exiftool"),
+        );
+
+        let mut cursor = 0;
+        let out = backfill_utc_dates_capped(&library, &exif_dao, &reel_dao, NY, &mut cursor, 10);
+        assert_eq!(out.converted, 1);
+        assert_eq!(out.unresolved, 1);
+
+        let fs_secs = std::fs::metadata(tmp.path().join("uv-long.mp4"))
+            .ok()
+            .and_then(|m| crate::utils::earliest_fs_time(&m))
+            .map(|t| chrono::DateTime::<chrono::Utc>::from(t).timestamp())
+            .expect("fs time");
+        let r = row(&conn, video);
+        assert_eq!(r.date_utc, Some(fs_secs));
+        assert_eq!(r.camera_offset_minutes, None);
+        assert_eq!(r.date_taken_tz_source.as_deref(), Some("utc_direct"));
+        // Legacy pair stays frozen.
+        assert_eq!(r.date_taken, Some(1_000_000_000));
+        assert_eq!(r.date_taken_source.as_deref(), Some("exiftool"));
+
+        // The read-error row keeps its legacy date and stays pending.
+        assert_eq!(row(&conn, junk).date_utc, None);
+        assert_eq!(remaining(&exif_dao, 1), 1);
+    }
+
     #[test]
     fn utc_backfill_unresolvable_rows_cannot_stall_the_rows_behind_them() {
         let (tmp, library, conn, exif_dao, _face_dao) = setup();

src/date_resolver.rs

@@ -48,7 +48,7 @@
 //! In-class files invert that rule deliberately (ADR-0001): a renamed
 //! screenshot or a chat-export timestamp is the only honest date left.
 
-use std::collections::HashMap;
+use std::collections::{HashMap, HashSet};
 use std::io::Write;
 use std::path::{Path, PathBuf};
 use std::process::{Command, Stdio};
@@ -259,7 +259,30 @@ pub fn resolve_dates_batch(
     paths: &[PathBuf],
     prior_exif_priors: &HashMap<PathBuf, ExifPrior>,
 ) -> HashMap<PathBuf, ResolvedDate> {
+    resolve_dates_batch_report(paths, prior_exif_priors).resolved
+}
+
+/// [`resolve_dates_batch`]'s result plus which files exiftool actually read.
+#[derive(Debug, Default)]
+pub struct BatchResolution {
+    pub resolved: HashMap<PathBuf, ResolvedDate>,
+    /// Paths exiftool returned an error-free entry for, whether or not it
+    /// held a usable date. A path in here that still resolved to `fs_time`
+    /// genuinely has no date tag (e.g. a muxer that wrote all-zero
+    /// `CreateDate`s); a path missing from it was never read — exiftool
+    /// absent, crashed, or reported an error for that file.
+    pub exiftool_read: HashSet<PathBuf>,
+}
+
+/// [`resolve_dates_batch`], also reporting which files exiftool read, so
+/// a caller can tell "no date tag in the file" from "exiftool didn't get
+/// a look at it" when the waterfall lands on `fs_time`.
+pub fn resolve_dates_batch_report(
+    paths: &[PathBuf],
+    prior_exif_priors: &HashMap<PathBuf, ExifPrior>,
+) -> BatchResolution {
     let mut out: HashMap<PathBuf, ResolvedDate> = HashMap::new();
+    let mut exiftool_read: HashSet<PathBuf> = HashSet::new();
     let mut needs_exiftool: Vec<&Path> = Vec::with_capacity(paths.len());
 
     for path in paths {
@@ -282,7 +305,10 @@ pub fn resolve_dates_batch(
         // One subprocess for the whole batch; each entry is already fully
         // resolved (legacy + new-contract fields).
         for (path, resolved) in exiftool_dates_batch(&needs_exiftool) {
-            out.insert(path, resolved);
+            if let Some(resolved) = resolved {
+                out.insert(path.clone(), resolved);
+            }
+            exiftool_read.insert(path);
         }
     }
 
@@ -310,7 +336,10 @@ pub fn resolve_dates_batch(
         }
     }
 
-    out
+    BatchResolution {
+        resolved: out,
+        exiftool_read,
+    }
 }
 
 /// Tag priority for exiftool extraction. First parseable value wins.
@@ -384,9 +413,12 @@ pub fn exiftool_available() -> bool {
 /// Append the exiftool arguments shared by the single-file and batch
 /// invocations: raw date strings (no `-d %s` — the resolver interprets
 /// them itself) with group prefixes for classification, plus the offset
-/// / GPS evidence tags for the cascade.
+/// / GPS evidence tags for the cascade. `-Error` puts exiftool's per-file
+/// read error (`ExifTool:Error`, e.g. "File format error") into the JSON
+/// — under `-q` it otherwise only reaches stderr, and the file's entry
+/// looks like a clean read that found no tags.
 fn exiftool_common_args(cmd: &mut Command) {
-    cmd.arg("-j").arg("-q").arg("-G1");
+    cmd.arg("-j").arg("-q").arg("-G1").arg("-Error");
     for tag in EXIFTOOL_DATE_TAGS {
         cmd.arg(format!("-{}", tag));
     }
@@ -430,7 +462,12 @@ fn exiftool_date_single(path: &Path) -> Option<ResolvedDate> {
 /// libraries with very long path components. Resolution is the same pure
 /// function as the single-file path, so the drain and ingest produce
 /// identical results for the same file.
-fn exiftool_dates_batch(paths: &[&Path]) -> HashMap<PathBuf, ResolvedDate> {
+///
+/// Keyed by every file exiftool read without error, `None` when the read
+/// found no usable date. Files it never reported (crash, partial output)
+/// or reported an error for are absent, unless the entry still carried a
+/// usable date.
+fn exiftool_dates_batch(paths: &[&Path]) -> HashMap<PathBuf, Option<ResolvedDate>> {
     let mut out = HashMap::new();
     if paths.is_empty() || !exiftool_available() {
         return out;
@@ -481,7 +518,8 @@ fn exiftool_dates_batch(paths: &[&Path]) -> HashMap<PathBuf, ResolvedDate> {
         );
     }
     for file in parse_exiftool_payload(&output.stdout) {
-        if let Some(resolved) = resolve_exiftool_payload(&file.date_tags, &file.evidence) {
+        let resolved = resolve_exiftool_payload(&file.date_tags, &file.evidence);
+        if resolved.is_some() || !file.read_error {
             out.insert(PathBuf::from(file.source_file), resolved);
         }
     }
@@ -810,6 +848,8 @@ struct ExiftoolFileData {
     source_file: String,
     date_tags: Vec<RawDateTag>,
     evidence: ExiftoolEvidence,
+    /// exiftool reported `ExifTool:Error` for this file (couldn't read it).
+    read_error: bool,
 }
 
 /// Split the exiftool `-j -G1` output into per-file date tags + evidence.
@@ -841,11 +881,16 @@ fn parse_exiftool_payload(stdout: &[u8]) -> Vec<ExiftoolFileData> {
             source_file: source_file.to_string(),
             date_tags: Vec::new(),
             evidence: ExiftoolEvidence::default(),
+            read_error: false,
         };
         for (key, value) in obj {
             let Some((group, tag)) = key.split_once(':') else {
                 continue;
             };
+            if group == "ExifTool" && tag == "Error" {
+                data.read_error = true;
+                continue;
+            }
             let Some(v) = value.as_str() else {
                 continue;
             };
@@ -962,6 +1007,29 @@ mod tests {
         assert!((lon - (-74.006)).abs() < 1e-9);
     }
 
+    /// A clean read whose date tags are all exiftool's zero placeholder is
+    /// "no usable date", not a read error; `ExifTool:Error` is the error
+    /// flag and isn't mistaken for a date tag.
+    #[test]
+    fn parse_payload_flags_read_errors_not_zero_dates() {
+        let json = br#"[{
+            "SourceFile": "/lib/zero.mp4",
+            "QuickTime:CreateDate": "0000:00:00 00:00:00",
+            "Track1:TrackCreateDate": "0000:00:00 00:00:00"
+        },
+        {
+            "SourceFile": "/lib/bad.jpg",
+            "ExifTool:Error": "File format error"
+        }]"#;
+        let files = parse_exiftool_payload(json);
+        assert_eq!(files.len(), 2);
+        assert!(!files[0].read_error);
+        assert_eq!(files[0].date_tags.len(), 2);
+        assert!(resolve_exiftool_payload(&files[0].date_tags, &files[0].evidence).is_none());
+        assert!(files[1].read_error);
+        assert!(files[1].date_tags.is_empty());
+    }
+
     /// #163: exiftool's default print conversion renders hemisphere refs as
     /// words. Before the fix "West" / "South" never matched, so western /
     /// southern coordinates came out positive and the GPS offset lookup hit