feat(shotload): shared shot-load state machine with a visible load chip

[obvious-autobuild]

Why

A cold shot load is silent: pick an uncached shot and the panels just sit there while the server fetches the bundle from FAIR-MAST. The audit found the compare flow solved this years before single-shot did — per-shot chips with queued/fetching/ready/error states, AbortController cancellation, Retry/Drop — but that machinery was private to the compare fan-out. This PR generalizes it into one shared shot-load machine and wires every shot fetch through it, closing the single-shot gap without changing the compare flow's behavior.

What

• web/shotload.js (new) — THE one shot-load state machine (spec §2, L4): per-shot queued → fetching → ready | error, AbortController supersede by generation, honest per-shot error reasons, listener notification. A superseded fetch lands back in queued — an abort is a supersede, not a failure.
• Single-shot loads run through the machine (web/replay.js loadShot) and render a load chip in Replay's status cluster: loading shot #30192… while fetching; on error, the reason plus working Retry and Drop. The silent cold-load is gone.
• Compare chips read the same machine — the hand-rolled cmp.st/cmp.err/cmp.ctrls maps are deleted; cmpFetch becomes a thin payload router over the machine. Cancel returns in-flight chips to queued instead of marking them "failed: fetch cancelled", matching the spec's state table.
• getReplay caches each promise with the signal it was created under, so a new-signal fetch never adopts an in-flight promise an older signal is about to abort; a failed fetch evicts only its own cache entry.

Scope: web/ only (new shotload.js, edits to replay.js, index.html script order, style.css chip styles). No backend changes, no build step, ES5-compatible IIFEs, tokens extended (--chip colors reuse the existing ok/warn/bad semantic tokens).

How to Review

• Read web/shotload.js first (~90 lines) — the whole state contract is there.
• The compare diff in replay.js is a mechanical map→machine swap; payload ownership (cmp.data, replayCache) is unchanged, so re-entering compare refetches exactly as before.
• data-guide anchors, window.FusionLab, and fusionlab:* events are untouched (compat contract L3).
• Known pre-existing observation (not from this diff): after a failed single load, navigating to compare keeps the stale failure line in the shot-state header until the reference resolves; the ref's renderShot wiring is byte-identical to the base (git diff cdb157d..HEAD -- web/replay.js). Flagging for the compare-workspace lane rather than patching out of scope here.

Acceptance criteria

• Machine states queued → fetching → ready | error with supersede and honest errors (web/shotload.js)
• Visible load chip in Replay's status cluster for single-shot loads; error chip carries reason + Retry · Drop
• Compare chips refactored onto the same machine (state maps removed)
• V6 evidence recorded against the tested head SHA: throttled load shows fetching → ready; failed fetch shows the error chip with working Retry (QA records tc-1, tc-2 below)
• Compare flow smoke-test unchanged (V2 slice — QA record tc-4)
• uv run pytest -q green (352 passed, 1 skipped), uv run ruff check clean, uv run mypy fusionlab clean (V1)
• Touches only web/ + no lessons.json copy changes needed; stored-energy units and naming untouched

Evidence (QA records against head 946687b8bbe6321b80a7e997868ded604c056369)

Recorded via obvious autobuild upload (15 s /replay throttle proxy; uncached-shot failure is shot #12017's real FAIR-MAST 502):

• tc-1 — V6 fetching → ready (fl_xVRldRRr, fl_sS18tvvd): status cluster chip "loading shot #30192…" while panels keep last-valid data; then the chip is gone with #30192 rendered.
• tc-2 — V6 error + Retry + Drop (fl_aCWLHiPh, fl_2wMHy7C2, fl_rE9smba0, fl_k4MKxGpp): error chip "couldn't load #12017 — shot 12017 could not be fetched from FAIR-MAST" with Retry · Drop; Retry restarts "loading shot #12017…" and honestly re-errors; Drop clears the chip and restores the picker.
• tc-4 — V2 slice compare smoke (fl_IOhwFA7E, fl_rkTA1DqG): deep-linked compare boots both chips "fetching from the archive…", then both ready with per-shot traces and the ΔW strip.

CI is the final gate; pytest re-runs via the pre-push hook.

Sibling PRs (Direction C release lane)

Lane	PR	Status
Router (prereq)	#58	✅ merged (cdb157d)
Explore workspace	#60	✅ merged (1b752ec)
Lane A: shot-load machine (this PR)	this PR	🔄 ready for review
Lane A: edit drawer	TBD	not opened
Lane A: compare workspace migration	TBD	not opened
Lane B: command palette	TBD	in progress by sibling worker
Promotion to main	#50 (draft)	after the parity audit

Human author: Tapan Ganatma Nakkina ([email protected])

🔗 Obvious Project · 🧵 Obvious Thread · 🔀 Obvious PR Artifact