The pluggable harvest layer

The Nominate ask promises that a no is a fine answer because "the kit will grow a pluggable harvest layer anyway." This plan is that promise costed out, so the fallback is an execution and not a scramble.

What actually touches Campaign Brain (measured, not assumed)

Six call sites, two concerns, one loose thread:

Concern	Where	Calls
Crawl	harvest main loop	list workspace jobs; get one job
	harvest dispatchDiscovery	dispatch a crawl
	signal-crawls script	dispatch a crawl
	rfp-watch --hunt	dispatch a crawl
Geocode	harvest geocode()	one cbgeo search
(loose thread)	manual curation	plain HTTPS fetch of result artifacts from cbfiles

cbemail is already off the list (the sending design moved transport to Proton). The geocache sits ABOVE the geocode call and caches successes forever, so a provider swap never re-spends lookups already made.

The seam: two interfaces, one file

scripts/housecalls-providers.mjs, exporting factories keyed by platform.json's new providers block (per-box choice, like everything else in that file):

{ "providers": { "crawl": "cbintel", "geocode": "cbgeo" } }

CrawlProvider (shapes are exactly what the pipeline consumes today):

listJobs()                → [job_id, ...]
getJob(id)                → { job_id, status, completed_at, result }
dispatch(query, opts)     → { job_id }
fetchArtifacts(job, dir)  → saves crawl_result / evaluation / pages, returns paths

fetchArtifacts is new and overdue: it formalizes the curl-the-result-url step that has been manual since the first crawl landed, CB-backed or not.

GeocodeProvider:

geocode(query) → { lon, lat } | null

The backends worth having

Crawl:

• cbintel: today's behavior, extracted verbatim. The native fit.
• manual (BUILT 2026-09-06): the honest degraded mode for a pilot without a workspace. dispatch files a request JSON in a directory that writes its own README; a human fulfills it by editing the file; the same loop harvests it like any platform result. Five end-to-end selftests against a real directory. Zero external dependencies; the curation discipline is identical.
• mock: fixture-driven, for the selftests. Worth building even on a Nominate yes: today the selftests cover the pure functions, and a mock provider lets the MAIN LOOP be tested end to end for the first time.
• Named for later, not built now: a self-hosted SearXNG + fetcher backend (fully anti-cloud discovery), and a session-Claude backend, which would also close the structuring-gap seam since the model that finds pages can emit prospect rows.

Geocode:

• cbgeo: today's behavior.
• census (BUILT 2026-09-06): better than the plan imagined. The Census geocoder API turned out to be address-only (city queries: no match), but the Census Gazetteer is vendorable: places + county subdivisions merged into lib/housecalls/places.json by scripts/vendor-census-places.sh (1,491 Michigan entries, incorporated places outranking townships on name collisions), making geocoding a local dictionary hit with NO network at harvest time. A/B against the live cache: agreement everywhere except Holland, which genuinely straddles the Ottawa/Allegan line (both answers are defensible), and Ada, which only the subdivisions file knows (hence the merge). An unknown city returns an honest null, never a guess.
• nominatim-public: OSM's public instance, 1 request/second usage policy. Fine at our volume with the forever-cache; a good third option.

What stays OUTSIDE the seam, by design

The rubric, the conservative extraction, the PII firewall, the stage ratchet, k-anonymity, the geocache format, and every rule on the page. Providers are dumb transports. A backend can change where pages and coordinates come from; it can never change what the machine is allowed to do with them.

Execution and cost

1. Extract the five call sites behind the two interfaces; cbintel and cbgeo backends are cut-and-paste of existing code (half a day, including re-running all six selftest suites).
2. mock provider + first main-loop selftests (2h, pays for itself).
3. manual crawl provider + census geocode provider (2h combined; the census API is one HTTPS GET).
4. Checklist and operator-config docs updated: phase 0's "platform access" line becomes "choose your providers," with cbintel as the recommended default where a workspace exists.

Total: about one day, same as the template extraction, and the two days compose: do this first and the template ships with the seam in it.

Triggers

• Nominate says not-now: this plan executes before the pilot weekend (the ask's after-the-answer section already promises exactly that).
• Nominate says yes: steps 1 and 2 still happen at leisure, because the mock provider improves the tests and the seam keeps the yes from ever becoming load-bearing dependence. Steps 3 and 4 wait until a pilot actually needs them.