feat/atlas-scan

seg_count double-counts on restart — upsert accumulates across runs

scripts/atlas/distribucion/dist_enumerate.py:102

The ON CONFLICT upsert uses `seg_count=COALESCE(seg_count,0)+excluded.seg_count`. There is no per-tile checkpoint, so a restart re-processes every tile and doubles segment counts for all circuits already written. `seg_count` is analysis-only (not billing), but it silently corrupts a key sizing proxy.

In-memory `enumerado` flag causes re-enumeration on restart, compounding seg_count inflation

scripts/atlas/distribucion/division_run.py:51

`enumerado = False` is reset on every process start. If killed after enumeration completes, the next run re-enumerates the full division, doubling seg_count for all circuits via the additive upsert. The other two phases (trafos, capacity) are genuinely resumable via DB-state checks; enumeration has no such guard.

`asign` keyed by substation name — name collisions merge distinct subs into one bucket

scripts/atlas/distribucion/dist_compare.py:94

Two distinct substations with the same name (plausible in CFE's dataset, e.g. two 'HERMOSILLO' nodes in different zones) would share one `asign` key, routing circuits of one to the other. Fix: key by coordinate tuple `(round(lat,4), round(lng,4))` instead of `s['n']`.

`porpunto` rounding may collapse physically distinct substations 11 m apart

scripts/atlas/distribucion/dist_compare.py:75

4 decimal places ≈ 11 m precision. Two distinct substations within 11 m (e.g. in a dense urban compound with slightly different stored coordinates) are merged, losing one's PF signals. Tighter rounding (5 decimal places ≈ 1.1 m) reduces false merges.

TLS certificate verification disabled (`CERT_NONE`) in dist_common.py

scripts/atlas/distribucion/dist_compare.py

Intentional — CFE's cert doesn't chain to Python's trust store. Acceptable for a research script reading public data. Risk escalates if this code graduates to a production ingestion path. Consider cert pinning as a middle ground.

Unclosed file handle on read path (json.load(open(PF)))

scripts/atlas/distribucion/dist_compare.py:70

CPython's GC closes it at scope end; no real risk. Wrap in `with` for hygiene.

File handles opened without context managers — write path risks truncated output on exception

scripts/atlas/distribucion/dist_compare.py:145

`json.dump(out, open(dst, 'w'), ...)` — if json.dump raises mid-write, the file is left truncated with no cleanup. Wrap in `with open(dst, 'w') as fh: json.dump(out, fh, ...)`.

`defaultdict` imported but used inconsistently — rest of file uses `setdefault`

scripts/atlas/distribucion/dist_compare.py:106

`porpunto` and `asign` use `dict.setdefault()`; only the `kv` accumulator inside the per-circuit loop uses `defaultdict(int)`. Remove the import and use `kv.get(k, 0) + 1` for consistency.

`dist_capacity.main(div, 3)` hardcodes worker count, not surfaced to operator

scripts/atlas/distribucion/division_run.py:73

The `3` is the thread worker count for the capacity phase — a meaningful performance knob. It is invisible in the CLI help and the docstring. Either document it or expose `--workers`.

`# (1b)` comment marker appears in code but not in the module docstring

scripts/atlas/distribucion/dist_compare.py:85

The docstring enumerates fixes as (1), (2), (3). The `# (1b)` marker in the code has no matching entry. Either add '1b' to the docstring or merge the comment into '# (1) continued'.

`str(e)[:60]` in exception handler drops the error class name

scripts/atlas/distribucion/division_run.py:79

Logs truncate to 60 chars of the message, losing the exception type. `f'{type(e).__name__}: {str(e)[:50]}'` gives faster diagnosis at zero cost.

No tests for the one-to-one circuit assignment invariant in dist_compare.py

scripts/atlas/distribucion/dist_compare.py

Fix 1 corrects a real double-counting bug (111 assignments over 71 circuits). A 3-line pytest asserting `len(assigned) == len(set(c['circuito'] for c in assigned))` would guard against regression. Low priority for a research script; high value if this becomes a recurring pipeline.

403 abort loses current tile's fetched data; restart replays entire tile loop from tile 1

scripts/atlas/distribucion/dist_enumerate.py:89

When a 403 is encountered, `raise` fires before the current tile's `acc` dict is committed. The tile's data is dropped. On restart, the tile loop begins at index 0, re-fetching all previously committed tiles. Quick fix: `con.commit()` before `raise`. For full resilience, store `last_tile_committed` in a `meta` table and slice `tl[last_tile:]` on re-entry.

`libre()` probes `subestacion` layer, not the layers actually being scraped

scripts/atlas/distribucion/division_run.py:31

The availability probe queries `sigdis:subestacion` but the active phases scrape `sigdis:linea_mta_hc` and `sigdis:banco_trans_dis`. Layer-specific rate-limiting would cause a false 'unblocked' signal, burning a cycle before hitting 403. Probe the layer matching the current phase.

`huerfanos` list uses dict-identity O(n×m) check — fragile if circuits are reloaded from DB

scripts/atlas/distribucion/dist_compare.py:129

`not any(c in v for v in asign.values())` relies on the same dict objects being in both `circuits` and `asign.values()`. Works now, but breaks silently if circuits are ever reloaded from SQLite. Fix: `asignados = {c['circuito'] for v in asign.values() for c in v}` then test by ID.

O(n×m) nearest-sub loop will not scale to full-division or national runs

scripts/atlas/distribucion/dist_compare.py:87

Pilot: 74 circuits × handful of subs = fine. Full NW division: potentially thousands × hundreds = ~10⁶ Haversine calls in pure Python. A spatial index (scipy cKDTree or bounding-box pre-filter) would give O(n log m) lookup. Not urgent now; note for national run.