fix: improve signal accuracy and optimize tweet ingestion

[Anniew31]

Changes:
Bug A — sentiment-analyzer.ts:18 — keywords 'will happen' and 'going to' were in BULLISH_KEYWORDS but the analyzer tokenizes text word-by-word. Can't be matched to so got rid of a word to make exactly one word long
Bug B — sentiment-analyzer.ts:56 — negation skips through modifiers. "not very bullish" → prevWord of "bullish" is "very" so the "not" two positions back was invisible and the phrase scored as bullish. Fixed to scan 2 words back, skipping modifiers when determining negation.
Bug C — signal-generator.ts:175 — HOLD threshold out of sync with bot cofig generateSuggestedAction returned HOLD for edge < 0.10, while the bot's
default BOT_MIN_EDGE=0.05 and computeUrgency labelled 0.05–0.10 as
"medium urgency" Lowered threshold to 0.05 to match.
Bug D — kalshi-client.ts:168 — all-time volume as 24h fallback
volume24h: km.volume_24h ?? km.volume ?? 0 fell back to km.volume, which
is Kalshi's total lifetime volume. Changed to use 0 when volume_24h is absent.
Bug E — api/cron/collect-tweets.ts — sequential KV writes
storeTweet was called with await inside a for...of loop, blocking on each write before processing the next tweet. Changed to collect AnalyzedTweet objects into an array first, then fire all writes in parallel with Promise.allSettled(tweetsToStore.map(storeTweet)).
Bug G — src/analysis/keyword-matcher.ts — emojis double-counted so a single emoji like 🚀 contributes 2 to emojiCount. That meant the threshold was triggered at 8 emojis. Fixed by replacing the regex with [...text].filter(c => (c.codePointAt(0) ?? 0) > 0xFFFF).length which made each emoji count as 1. Also Adjusted the threshold to 8 to keep behavior same

[vercel]

@Anniew31 is attempting to deploy a commit to the Victor's projects Team on Vercel.

A member of the Team first needs to authorize it.