feat(#203-A3): ingestor writes file-bearing assets under ds_<hex>

[LukasWodka]

#203 Track A3 — RFC-0003 D9 phase 2 / D16

Today's #255 (scoped mount) and #569 (engine path) both re-keyed dataset access onto the physical ds_<hex> handle. But the ingestor still hard-refused file-bearing categories under PER_INGESTION_TABLES, because their assets landed in the shared label tree (STORAGE_PATH/<label>) where a same-label re-ingest would overwrite an earlier dataset's files. This PR moves the file tree onto the same handle and lifts the refusal.

Change

• config.py — DEST_PATH now keys on a new DEST_TABLE (defaults to the TABLE_NAME label, injectable via set_dest_table). This is the one knob: every file-copy primitive, the text profiler, the duplicate validator, and the source-reclaim path already read DEST_PATH off the run config, so they all follow automatically.
• ingestors/base.py — when per_ingestion_tables, inject the ds_<hex> handle: config.set_dest_table(physical_table_name). File-bearing assets now land in STORAGE_PATH/ds_<hex>/, matching the scoped mount (refactor(P3b): move validator factories into the ModalityRegistry #255) and get_dataset_path (#569). The file-bearing refusal is removed.
• tests — flipped the "refuses file-bearing" test to assert the handle injection; added a real-Config DEST_PATH test. Flag off is byte-for-byte unchanged (DEST_TABLE = label, set_dest_table never called).

Why it matters

Unlocks image / object-detection / semantic-segmentation / keypoint datasets (the bulk of tracebloc's use-cases) under per-ingestion isolation, and completes the D9-phase-2 file half that #255 + #569 were built to enable. It's the ingestor change the on-HOLD client#490 v0.8.0 image is waiting on.

Scope note

The file table lock intentionally stays keyed on the label — that's pre-existing #408 behavior (the row store already moved to ds_<hex> while the lock stayed on the label); it's a concurrency guard, not a data-isolation unit, so it's out of scope here.

Verified

Full ingestor suite 1903 passed, 1 xfailed; black clean.

Targets develop.

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes on-disk layout for file-bearing datasets when PER_INGESTION_TABLES is on; misalignment with scoped mounts or backend path resolution could strand or mis-associate assets, though flag-off behavior is unchanged.

Overview
With PER_INGESTION_TABLES enabled, file-bearing ingests no longer error at construction; instead the run repoints the on-disk file tree from the shared label directory to the same ds_<hex> handle used for the MySQL row store.

Config introduces DEST_TABLE (defaults to the user label) and set_dest_table, so DEST_PATH becomes STORAGE_PATH/<DEST_TABLE>. BaseIngestor calls set_dest_table(physical_table_name) when the flag is on, which pulls file copy, text profiling, duplicate checks, and source reclaim along without separate changes. TABLE_NAME and flag-off behavior stay as today.

Tests now assert handle injection for image-classification under the flag and DEST_PATH repointing; package version 0.8.0 → 0.8.1.

^{Reviewed by Cursor Bugbot for commit 110d3d8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[LukasWodka]

bugbot run

[LukasWodka]

bugbot run

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 110d3d8. Configure here.}

[saqlainsyed007]

Verdict: APPROVE.

Reviewed the full-file context, not just the hunks. The change is correct and consistent with the row-store side of RFC-0003 D9 phase 2.

What I verified:

• DEST_PATH repoint — DEST_TABLE defaults to TABLE_NAME, so flag-off is byte-for-byte unchanged; set_dest_table(physical_table_name) is the only writer, and every consumer (file_transfer copy primitives, text_profile, duplicate_validator, reclaim_source) reads DEST_PATH off the same run config, so the single knob genuinely propagates.
• Path safety — physical_table_name is always ds_<uuid4().hex> (hex-only), so no traversal/injection surface is introduced by moving the file tree.
• Reclaim — _reclaim_source still keys the source gate on TABLE_NAME (the CLI's .tracebloc-staging/<label> dir) and only uses DEST_PATH for the overlap backstop; staging (<label>) and dest (ds_<hex>) don't overlap, so reclaim still fires correctly.
• Table lock — stays on the label via TableLock(self.table_name, …), matching the scope note; concurrency-only, not a data-isolation unit.
• Refusal removal — clean: _FILE_BEARING_CATEGORIES is still used (base.py:547, 1031), no dead import.
• Version bump — 0.8.0→0.8.1 is single-sourced in __init__.py and parsed by setup.py:_read_version; nothing else to update.
• Tests — real assertions (handle injection via set_dest_table.assert_called_once_with, and a real-Config DEST_PATH repoint asserting label untouched), not vacuous.

Non-blocking observation (no change requested): set_dest_table permanently mutates the shared Config override with no reset path. This is safe because cli/run.py builds one Config → one ingestor per process; if a future caller ever reused one Config across a flag-on then flag-off ingestor, the stale ds_<hex> DEST_TABLE would leak into the flag-off run. Worth a one-line comment noting the single-run-per-config invariant, but not reachable today.

[LukasWodka]

/fr-pass