feat(kick): fix campaign discoverability — open-campaign directory fallback + shell rows (#60)

[steethebutcher]

Fixes the Kick half of #60 ("even if I add kick campaigns like WoW Forever it won't mine it, and I must add whitelist channel like rewardstation to mine the wow drop").

Root causes (two independent defects)

1. Open campaigns had no channel candidates. Most Kick campaigns are open (channels: [] — accrue on any live channel in the category). ListEligibleChannels only considered the campaign's own channels, a category pool built exclusively from sibling campaigns that embed channels, and manual operator channels. WoW Forever is the only WoW campaign, so the pool was empty → "no eligible streams live" forever → unmineable without a manual channel registration.

2. Non-whitelisted campaigns were invisible (deadlock). The backend dropped them (GameFilter continue) and the scraper re-dropped them, so they never reached the campaigns table or the /drops Discoverable tab. A user couldn't see "WoW Forever" to whitelist it — the only way in was adding the game blind on /priority. Twitch deliberately avoids this deadlock with shell rows; Kick now does too.

Commits

1. feat(kick): directory fallback for open campaigns with no channel pool

When an open campaign's pooled candidates yield nothing live, consult the public category directory, then merge + probe:

• Parses the campaign payload's authoritative game_slug (previously discarded) so the directory is addressed correctly instead of re-deriving the slug from the display name (cf. the Twitch slug bugs in "watcher no eligible streams live" logged despite active Twitch streams with drops for campaigns #61).
• Directory results are viewer-sorted and merged into categoryChannels so later picks work even if the feed goes away.
• Every candidate still passes probeLive's live + on-category verification — a wrong slug can only produce zero results, never the junk picks that got the old generic feed removed.
• Restricted campaigns (own channels[]) never consult the directory.

2. feat(kick): emit shell rows for non-whitelisted campaigns

GameFilter becomes a should-map-benefits gate (Twitch parity, campaigns.go:236): non-whitelisted campaigns keep ID/Game/Name/Status/window/AllowedChannels but carry empty Benefits and skip channel-pool caching. The scraper passes shells through to the persister so they surface in Discoverable + the game picker. Safety is the same mechanism that already makes Twitch shells safe: the watcher's own AllowGame + never-mine-empty-Benefits.

3. fix(kick): directory feed is global + paginated — filter client-side

I live-checked the public /stream/livestreams/{slug} endpoint before wiring it in (2026-10-03) and found the dead-code assumptions wrong:

• The feed ignores the category slug — rust vs world-of-warcraft return the identical global list (same failure family that got the old feed removed).
• It paginates (per_page=5, ?page=N), and each entry carries its own categories[] (name+slug) + channel.slug.

So DiscoverChannelsForCategory now walks ≤4 pages and filters client-side by slug or display name (case/space-insensitive — the real feed mixes case, e.g. "Project-Zomboid"). Live-verified by simulation against the production feed: a 4-page scan extracted categories correctly and the filter matched channels for a category present in the window. Combined with probeLive's independent verification, junk picks remain impossible whether or not Kick restores per-category filtering.

Known limitations (deliberate)

• Directory discovery is progressive, not exhaustive. The window is 4 pages × 5 entries of a global feed; a category absent from this cycle's window gets picked up on a later scan as the feed shifts. Fine for mining cadence (retried every pick cycle), documented here so nobody expects instant results for niche categories.
• /api/v1/drops/campaigns has no pagination handling — if Kick ever pages that endpoint, campaigns would silently truncate. Unverifiable without a live authed response; flagging for awareness.
• No new config knobs — the fallback is automatic and self-verifying.

Verification

• go build ./..., go vet ./... clean; full test suite green (24 packages).
• 12 new/updated tests: directory happy path (with off-category entry filtered pre-probe), payload-slug precedence, wrong-category rejection at both layers, restricted-campaign exclusion, pool persistence across picks, shell-row contract ×2, scraper pass-through.
• Directory endpoint shape live-verified via curl + parsing simulation (details in commit 3).
• Still needs a live mining verification on a deployment with a Kick account: WoW Forever should appear in Discoverable, whitelist it, and mine without manual channel registration — happy to run that before merge if you want (same gate the Spade PR used).

Out of scope

Twitch slug derivation (#61 — JourneyOver's fixes in flight), device-code login (#48), pipeline redesign (#50).