docs(Tanzania): back-casting is a country fact, not a food_acquired quirk (+ dispatcher rule)

[ligon]

Two changes, one cause.

1. The Tanzania NPS back-casting is documented as a country-level fact

Tanzania handles split-offs differently from most — possibly all — other surveys here. When a child leaves and founds a household, UPD4 creates new observations in the earlier waves too, carrying the ancestor's record back under the descendant's own panel identifier.

This was already written up — but only inside the #114 entry, in the context of food_acquired. That framing made it read as a food-specific quirk. It is a property of the panel construction, and every 2008-15 table inherits a choice from it.

New top-level * Survey design section records:

variable	label
UPHI	Uniform Panel Household Identifier at NPSY4
r_hhid	NPS HHID: y1_hhid, y2_hhid, y3_hhid, y4_hhid

round	rows	distinct UPHI	distinct r_hhid
1	20,039	5,594	2,956
2	155,097	8,163	3,924
3	189,962	9,998	5,010
4	84,337	4,961	4,341

In round 1, 1,526 r_hhid carry more than one UPHI, up to 11.

The load-bearing consequence, which is why this needed writing down: neither key produces a signal the usual duplicate-index instruments can see. Keyed on UPHI the index is unique, so nothing warns — absence of duplicates is not evidence the key is sound. Keyed on r_hhid the back-cast copies collapse, and since their payloads are identical the collapse is value-lossless and equally silent.

It also records that the tables are currently inconsistent — food_acquired de-duplicates to r_hhid (408,707 → 228,281); 2008-15 shocks keys on UPHI — and that resolving this is a maintainer grain decision, not something to change as a side effect of unrelated work. The #114 entry gains a scope note pointing up at the new section.

2. A dispatcher rule in CLAUDE.md

Any agent scoped to a country must be told to read countries/{C}/_/CONTENTS.org first.

The add-feature skill already says "Read existing documentation FIRST … Start here". The gap is that a dispatched agent reads only what its brief tells it to — so this is the dispatcher's responsibility, and that is where the rule now lives.

Three things it adds beyond "read the file":

• read the LOGBOOK/WAITING/TODO entries too — a closed GitHub issue can still have a live caveat parked there (#114 is closed; its entry is still WAITING);
• a finding that contradicts CONTENTS.org is a result to report and quote, not licence to assume the file is stale;
• add any genuine idiosyncrasy found that is not recorded.

The cost that motivated it

An agent reviewing Tanzania's .first() sites concluded shocks had a broken key because its i values shared nothing with household_roster. I endorsed that, then flagged it as a regression, then withdrew both — because Tanzania/_/CONTENTS.org already documented the whole phenomenon. Two rounds of confident, wrong framing from skipping one file.

Docs only. No behaviour change; the UPHI/r_hhid question is deliberately left open (see PR #651).