Fix the inverted EDUCQUAL to highest_education mapping

[juaristi22]

Fixes #459.

The EDUCQUAL_MAP read the FRS highest-qualification codeframe upside down: code 1 was mapped to NOT_COMPLETED_PRIMARY and codes 17–21 to TERTIARY, but the frame starts at 1 = Doctorate and descends. highest_education feeds live model logic (maintenance loan system assignment, DSA course eligibility, student loan plan assignment).

This cherry-picks the existing correction from feat/impute-below-threshold-borrowers (8f95629, authorship preserved) onto main, now verified two independent ways:

• The official data dictionary: FRS 2023-24, UK Data Service SN 9367, DOI 10.5255/UKDA-SN-9367-2, adult table — every label in the corrected map matches verbatim (1 = Doctorate or MPhil, 17 = A-Level or equivalent, 18/19 = Welsh/Scottish Baccalaureate, 36–49 the GCSE band, 83–85 entry level). Code 87, carried by ~13% of adults, has no value label in the dictionary and falls to the UPPER_SECONDARY default.
• Disclosure-safe aggregates from the 2023-24 tab (28,590 adults): code 1 is 1.8% of adults with the highest mean weekly earnings (~£744); codes 18/19 have fewer than 10 respondents each; the GCSE band carries the population mass. Under the inverted map the weighted TERTIARY share came out near 10%; the corrected frame gives ~28%.

Under the corrected map the weighted highest_education composition of the base FRS moves from {TERTIARY 9.9%, POST_SECONDARY 21.3%} to {TERTIARY 28.1%, POST_SECONDARY 8.4%} — expect downstream shifts in student-support and education-conditioned outputs on the next data build.

The microcosm migration adopted the same corrected frame as a signed difference in PolicyEngine/microcosm#703, so landing this keeps the two pipelines aligned.

Test suite: 484 passed, 107 skipped (licensed-data tests skip without staged inputs), no failures.

🤖 Generated with Claude Code

[vahid-ahmadi]

Verified against both sides of its provenance: the corrected map is identical, code for code, to the frame the microcosm migration adjudicated and verified in PolicyEngine/microcosm#703 — where I checked it against the harmonised codeframe during that review — and the per-code labels here match the SN 9367 dictionary labels quoted in the body (1 = Doctorate or MPhil, 17 = A-Level or equivalent, 18/19 the near-empty baccalaureates, 36–49 the GCSE band, 83–85 entry level). The two-way evidence — dictionary labels verbatim plus the disclosure-safe aggregates (code 1 at 1.8% of adults with top mean earnings is not "not completed primary") — is exactly the standard a coded-domain correction should meet, and enumerating every code with its label instead of the old range-loops means the next dictionary revision gets caught by a diff rather than absorbed by a for code in range(...).

The composition move ({TERTIARY 9.9%, POST_SECONDARY 21.3%} → {28.1%, 8.4%}) lands the base FRS at a plausible UK degree-holder share for the first time, and flagging the expected downstream shifts in student-support outputs up front — rather than letting the next data build discover them — is the right disclosure. Keeping the two pipelines on the same frame is itself worth landing this promptly, since every microcosm head-to-head from E3 onward diffs against this repo's output.

Two small things:

1. The new comment header says "Based on FRS 2022-23 documentation (SPSS value labels)" while the changelog and PR body cite the FRS 2023-24 dictionary (SN 9367). If the map was verified against 2023-24 — which the body's evidence says it was — the comment should name that vintage (or both, if the 2022-23 labels were checked too). After the migration's region-map lesson, the vintage a coded-domain map cites is load-bearing documentation.
2. Code 87's ~13% of adults defaulting to UPPER_SECONDARY is preserved behavior and correctly out of scope here, but it's now the single largest source of highest_education uncertainty — an eighth of adults assigned by fallback. Worth a follow-up issue so it doesn't hide behind the fixed map: even a crude imputation from age/earnings would beat a constant default for a bucket that size.

With the vintage comment fixed, this is ready to land as far as I'm concerned.