feat(sequences): re-run cone matching when relabeling back to wildfire_smoke

[MateoLostanlen]

• Context: when a sequence is mis-labeled as non-wildfire (e.g. other_smoke), it's detached from its triangulating alert and a fresh single-sequence alert is created. Reverting the label to wildfire_smoke previously left the sequence stuck in that lonely alert, breaking triangulation. This PR makes the revert path drop the lonely alert and re-run cone matching so the sequence merges back with its overlapping neighbours.
• Anchor the candidate-sequence window on the sequence's own last_seen_at (not utcnow()) so relabels done hours later still find the original neighbours within SEQUENCE_RELAXATION_SECONDS.
• Rename _attach_sequence_to_alert → attach_sequence_to_alert and extract _detach_sequence_from_alerts to deduplicate the detach/refresh block.

[codecov]

Codecov Report

❌ Patch coverage is 95.23810% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 92.31%. Comparing base (00743b2) to head (c5b1820).

Files with missing lines	Patch %	Lines
src/app/api/api_v1/endpoints/sequences.py	94.11%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #585      +/-   ##
==========================================
+ Coverage   92.30%   92.31%   +0.01%     
==========================================
  Files          56       56              
  Lines        2585     2603      +18     
==========================================
+ Hits         2386     2403      +17     
- Misses        199      200       +1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Acruve15]

It's an interesting PR, thx! I can do a thorough review by Monday if it's ok for you.

I want to make sure I understand what happens with the triangulation and how alerts/sequences/detections concepts are linked :)

[Acruve15]

@MateoLostanlen I'll review by EOD

[Acruve15]

Hey @MateoLostanlen, I added some comments, the most important one on how to handle orphan alerts :)

Can be helpful, here is the current flow for context:

sequenceDiagram
    autonumber
    participant U as User
    participant L as label_sequence
    participant CM as cone matching
    participant A as AlertSequence

    Note over U,A: T0: Alert#100 triangulates {seq_B, seq_C}

    U->>L: label seq_B OTHER_SMOKE
    L->>A: detach seq_B from Alert#100
    L->>A: create lonely Alert#101 = {seq_B}
    Note over A: Alert#100 = {seq_C}, Alert#101 = {seq_B}

    Note over U,A: T1 (later): user reconsiders
    U->>L: label seq_B WILDFIRE_SMOKE

    rect rgb(255, 235, 235)
    Note over L,A: pre-PR: no-op, seq_B stuck in Alert#101 forever
    end

    rect rgb(235, 250, 235)
    L->>A: detach seq_B from Alert#101 (deleted, empty)
    L->>CM: anchor on seq_B.last_seen_at
    alt overlap found
        CM-->>A: link seq_B to Alert#100
        Note over A: Alert#100 = {seq_B, seq_C} again
    else no overlap (orphan risk)
        CM-->>A: returns None, no link created
        Note over A: seq_B has zero alert rows
    end
    end

Loading

Longer term: I was thinking we could soft-delete AlertSequence rows (with unlinked_at timestamps) would give us audit history for free and is useful when debugging "why did this alert lose triangulation?"?

[Acruve15]

For consistency?

Suggested change

	updated, camera, cameras, sequences, alerts, reference_time=sequence.last_seen_at
	updated, camera, cameras, sequences, alerts, reference_time=updated.last_seen_at

[Acruve15]

if this returns None (eg no neighbour in the window or cones don't overlap?) , the sequence has zero alert links. I suggest adding a fallback to make sure we don't have orphan alters.

alert_id = await attach_sequence_to_alert(
    updated, camera, cameras, sequences, alerts, reference_time=sequence.last_seen_at
)
if alert_id is None:
    new_alert = await alerts.create(
        AlertCreate(
            organization_id=camera.organization_id,
            started_at=sequence.started_at,
            last_seen_at=sequence.last_seen_at,
            lat=None,
            lon=None,
        )
    )
    session.add(AlertSequence(alert_id=new_alert.id, sequence_id=sequence_id))
    await session.commit()

If orphaning is the intended behaviour, a one-line comment is enough for me.