Feat/fasta prep backend

[t03i]

TLDR

MVP backend implementation.

Description

Architecture

[Browser]
   │ POST /api/prepare (multipart FASTA)
   │ GET  /api/prepare/:id/events  (SSE)
   │ GET  /api/prepare/:id/bundle  (download)
   ▼
[Caddy reverse proxy] ── existing TLS termination, SSE-friendly by default
   │
   ▼
[protspace-prep container]
   │ FastAPI app (uvicorn, single process)
   │ ├─ in-memory job registry: dict[job_id → JobState]
   │ ├─ asyncio.Semaphore(5) bounding concurrent active jobs
   │ ├─ asyncio task per job; pipeline runs via asyncio.to_thread
   │ └─ /var/lib/protspace-prep/jobs/<job_id>/ for bundle artifacts
   │
   ▼
[Biocentral API] (HTTPS, anonymous today)

SLO:

• 5 concurrent users supported
• resutls of 1k seqs in ~1 min

Closes: #236

[t03i]

1 min is currently absolutely unrealistic. For 840 sequences, we're breaching 3:30 likely because of Biocentral. Some more profiling is required.

[t03i]

Changed e2e to only run nightly/with label.
Changes regarding colab and progress already in now @tsenoner.

[tsenoner]

Review notes

Likely bugs

• app/src/explore/runtime.ts:140 — the embedding/annotating branch hardcodes 'Embedding sequences (~3 min)…', overriding the dynamic embeddingLabel from formatEmbeddingLabel(seqCount). A 50-seq job will show
  "~3 min" instead of "~20 sec". Should reuse embeddingLabel like the queued branch does.
• api.py:134 — mark_consumed(job_id) runs before the response body is sent; the bundle deletion is a
  BackgroundTask. If the SPA's download fetch fails mid-stream, consumed=True and the file is gone, so
  retry returns 410. Consider deferring mark_consumed until the BackgroundTask completes successfully.

Colab fallback UX

• We want the Colab notebook
  (https://colab.research.google.com/github/tsenoner/protspace/blob/main/notebooks/ProtSpace_Preparation.ipynb)
  to be a visible alternative — not only during the prep flow as the "Got a larger dataset?" note. Consider
  surfacing it on the explore page itself (e.g. near the FASTA upload affordance) so users know they can prepare
  the data themselves.
• When Biocentral is unavailable, the backend already returns BIOCENTRAL_UNAVAILABLE with a friendly message
  in _BIOCENTRAL_FRIENDLY_MESSAGE, but the message is plain text — the frontend currently doesn't render it
  as a clickable link to the Colab notebook. When the SPA receives an error with code === 'BIOCENTRAL_UNAVAILABLE', it should show a dedicated banner/notification with the Colab link and an
  explanation that the user can run the same pipeline there. Right now the user just sees an error toast.

Dead / inconsistent config

• Settings.biocentral_endpoint (config.py:19, PREP_BIOCENTRAL_ENDPOINT) is loaded but never read. Either
  pipe it into the protspace embed subprocess or drop the field.
• PREP_SEQUENCE_MIN_COUNT defaults to 1 in config.py, is overridden to 20 in docker-compose.yml, and
  isn't mentioned in services/protspace-prep/README.md. Pick one and document it.

Worth a TODO, not a blocker

• In-memory JobRegistry — container restart orphans live jobs. Job dirs survive on the volume but _jobs is
  empty, so /bundle returns 404 until the sweeper runs. Worth flagging in the README and deciding how the SPA
  should react.
• _classify_failure is only applied inside the embed/annotate TaskGroup catch (pipeline.py:115). A
  failure in project or bundle won't get the BIOCENTRAL_UNAVAILABLE reclassification. Not a real risk
  today, but a sharp edge if those steps ever go network-bound.

Smaller

• _looks_like_nucleotide is all(ch in ACGTUN) over ≥50 residues — theoretically can false-positive on
  pathological sequences (e.g. poly-A peptides). Consider a ratio-based check if false positives ever appear.