Pausing mid-edit no longer discards the Cluster Strength

[lstein]

Rebased onto master after #370 landed, and re-scoped: #370 changed what an empty field means, which changes what this bug is.

The bug

An empty Cluster Strength field is now a deliberate signal — it means "go back to a derived strength". The trap is that <input type="number"> reports an empty value for anything it cannot parse yet:

"0.4" -> value="0.4"
"0."  -> value=""      <- indistinguishable from a cleared field
"-"   -> value=""
"1e"  -> value=""

So an empty field alone cannot be read as the user asking for a derived value. Pausing for a second while retyping looks exactly like it, and throws away the number the album was tuned to:

album tuned to 0.4; user selects the field and retypes
  "0"   -> parseable, save armed
  "0."  -> value becomes "", i.e. "clear it"
  ...user hesitates for one second...
  -> POST /set_umap_eps/ {eps: null}
  -> the album is back on a derived strength, "auto" badge and all

The existing guard cannot catch this. if (eps !== null && Number.isNaN(eps)) is dead code for a number input: the sanitized value is "", so readSpinnerEps returns null, never NaN.

The fix

validity.badInput is the discriminator. The browser sets it only while the input holds text it could not turn into a number, so it separates "the user cleared this" from "the browser cannot parse this yet" — which reading .value alone cannot do.

Two smaller things in the same handler:

• The pending save is dropped before every early return, not after them. Otherwise a save armed by an earlier keystroke still fires a second later carrying a number the field no longer shows.
• A value below the spinner's own min is not saved. The server floors anything under MIN_CLUSTER_EPS, so the map ends up clustered at something other than what the spinner and the cluster-info modal report. (The ceiling is deliberately not enforced: a derived strength for a small album can legitimately exceed max, and refusing those would leave exactly those albums untunable.)

What the review passes added

Refusing to arm a save for half-typed text removed the only evidence that an edit was in progress — the debounce handle, which the albumIndexUpdated listener reads before re-resolving. Three passes of adversarial review (each finding reproduced before being fixed, each fix pinned by a test that fails without it) turned that into a guard that actually holds the field:

• epsEditPending / epsEditSeq / an in-flight save count. The field is held while the user is part-way through it, while a save it armed is in the air, and against a late get_umap_eps reply the user has since typed past — a derive can take a minute on a large album, long enough to change your mind, and the reply used to land on top of the number you had just stored, "auto" badge and all.
• Blur does not release the hold while unparseable text is still showing. Blur fires on a click anywhere else and on an alt-tab, and the half-typed text stays in the field. The hold ends when the text does.
• A failed save is no longer indistinguishable from a successful one. Neither the HTTP status nor a rejection was checked, so the map was redrawn at a strength the album does not have — most often against the 403 from require_no_lock, i.e. while the album is being indexed, which is exactly when someone is tuning it.
• A refused value is no longer sent either. Typing 0.005 armed no save, but the next redraw still asked the map for it and the server floored it — the exact divergence refusing to store it prevents.
• A save landing on a closed window rendered Plotly into a hidden container and consumed the dataChanged flag, so re-opening showed a zero-size plot and refetched nothing.
• A saved strength belongs to the album it was typed in. The save read state.album when the timer fired, so switching albums inside the one-second window silently wrote the number into the other album's config. (Pre-existing; fixed here because it is this handler.)
• A stored strength the map cannot cluster with is marked rather than silently presented as fine — versions before this one could store 0, -2 or 0.005, and the config file is hand-editable.

Tests

tests/frontend/umap-eps-debounce.test.js (new, 17 tests) and tests/frontend/umap-reindex-refresh.test.js (+5). 615 frontend tests pass, 728 backend, eslint/prettier/ruff clean.

The debounce suite runs on fake timers: five real 1.15s waits cost more wall clock than the whole rest of the frontend suite.

A new backend test pins the spinner's min attribute to MIN_CLUSTER_EPS — the frontend reads its floor off the element rather than hardcoding it, so the two could otherwise drift silently.

Caveat on the main one, now closed. jsdom sanitizes an unparseable value to "" but never sets badInput (verified — see typeUnparseable, which stubs it). These tests pin this module's logic; that a browser really sets badInput for a half-typed number is a platform guarantee, not something this suite proves. Confirmed by hand in a real browser — pausing mid-number no longer discards the album's strength.

Known, not fixed here

Three things the review turned up that are not this branch's to fix, now filed:

• /umap_data never reports the eps it clustered with, so a tuned large album shows a strength the map never used #380 — _shrink_eps_to_pair_budget shrinks large typed values too, so a 40k-image album saved at 2.0 clusters at 1.4 while the spinner, /get_umap_eps and the cluster-info modal all report 2.0. No client-side bound can prevent it; the server needs to report the eps it used.
• Cluster-info modal reports the spinner's value, not the strength the map was drawn with #381 — the cluster-info modal reads the raw spinner rather than the eps the current draw used, so it can report a third number again.
• Non-finite cluster_eps either 500s or wedges a worker thread forever #379 — the backend accepts non-finite cluster_eps: nan reaches DBSCAN (500), and inf makes _shrink_eps_to_pair_budget loop forever on any album past ~7k images, wedging a thread-pool worker. Not reachable by typing, but reachable by URL or a hand-edited config. Belongs with Refuse a Cluster Strength that DBSCAN cannot run with #375.

🤖 Generated with Claude Code