feat(grammar): allow bare !post — narrative from device metadata (#124)

[brockamer]

Summary

• Bare !post (no caption) now parses and produces a journal post — previously returned undefined and replied "Unknown command. Try !help", wasting a satellite send.
• The LLM builds the narrative purely from enrichment context (place name, weather, time) under a no-note system prompt that explicitly forbids inventing activities, feelings, or specifics not present in metadata. Observational third-person voice instead of first-person.
• note is now string | undefined on the post ParsedCommand variant. Whitespace-only notes are treated as bare (cleaner than the empty-string ambiguity).
• Closes Phase 2.5's last open issue — milestone goes 0/3 once this lands.

Surfaced 2026-04-27 during #111 production traffic turn-on (operator typed bare !post expecting "TrailScribe figures it out").

Acceptance criteria (from issue body)

• src/core/grammar.ts — bare !post parses as { type: "post" } (note omitted at the type level — distinguishes "bare" from "empty after trim")
• src/core/types.ts — note?: string on the post variant
• src/core/commands/post.ts — branches on cmd.note === undefined; defensive empty-narrative fallback (Posted from {placeName} or Posted from the field)
• src/core/narrative.ts — SYSTEM_PROMPT split into WITH_NOTE / NO_NOTE; buildUserPrompt omits the Note: line when absent
• Vitest cases — 1 grammar test (replaces the negative case), 3 narrative-prompt tests, 2 end-to-end pipeline tests (!post with GPS, bare + no-GPS edge case)
• docs/field-commands.md documents the bare mode
• !help reply text updates: !post [note] — blog

Tests: 332/332 (was 327; +5 new). Typecheck + lint clean.

Cost envelope (open question from issue body)

The risk note flagged whether bare-mode input tokens stay under PRD §6's <$0.05/transaction. Plan: probe staging with one bare !post + one with-note !post, compare ledger deltas. Will post the numbers as a PR comment before requesting merge.

Test plan

• Deploy to staging via pnpm deploy:staging (pre-approved, dry-run first)
• Probe bare !post against staging Worker; capture token usage from ledger
• Probe with-note !post against staging; compare deltas
• Confirm bare path stays under <$0.05/transaction; if not, investigate prompt-trimming before merge
• Real-device verification deferred until P2-16 (operator + Mini in field)

🤖 Generated with Claude Code

[brockamer]

Staging cost probe — done

Flipped staging IPC_INBOUND_DRY_RUN to true, redeployed (version b6be26e6), fired one bare !post and one with-note !post against the staging Worker, then read the ledger:2026-04 rollup before/after. Reverted IPC_INBOUND_DRY_RUN to false and redeployed (version 76591947) — staging is back to its original config; wrangler.toml has no diff.

Ledger delta (2 calls combined: 1 bare + 1 with-note)

Metric	Pre	Post	Δ
post requests	50	52	+2
prompt_tokens	22689	23621	+932
completion_tokens	13417	13667	+250
post usd_cost	$0.15664170	$0.16318770	+$0.00654600

Per-transaction mean: ~$0.003. That's 6.5% of the PRD §6 $0.05 envelope — well inside target. Both calls returned 200; bare took ~26s, with-note took ~12s (the bare path's longer latency is the model elaborating from sparse metadata vs. anchoring on a caption).

The probe doesn't separate per-call tokens (the ledger rollup is an aggregate), but the combined-mean already answers the open question: bare-mode cost is comfortably inside envelope. If the with-note path's typical token count (~240 prompt / ~180 completion in unit tests) is the baseline, the bare path's share of the +932/+250 delta is at most ~750/+250 = roughly 1.5× the with-note prompt count — which is the expected direction (model has more enrichment to lean on, not less) and still nowhere near the budget ceiling.

P2-16 real-device verification will confirm field UX (does the LLM produce a readable observational narrative with no caption?), but cost is no longer a merge blocker.

Ready for your merge approval.