XERK-427: give the raw-cursor loop's entry lookups their own budget term

[xerhab]

What

rawCursors in turma/archive.js computes per-file byte cursors for the raw archive layer each heartbeat. Its per-beat loop charges a budget for two kinds of work — one unit per manifest entry (the SELECT filePath row lookup) and one per file offered — but the budget was a single ARCHIVE_RAW_CURSOR_MAX (2000), which mirrors only the agent's file cap ARCHIVE_RAW_MANIFEST_FILES_MAX (2000).

So an in-cap agent offering F files across N transcripts cost the hub N + F work and overran by exactly N. The last N offers got no cursor; the agent read that as 0 and re-pushed from the start, and ingestRaw refused each on its offset check — a wasted small POST per affected file per beat. The hub's own comment claimed an in-cap agent "never reaches this"; it was not arithmetically true.

The interaction that made it worth fixing now (from the ticket): XERK-424's reserved backlog slice exists so older, unarchived sessions get their sidecars shipped — and because the manifest is newest-first, this budget truncated exactly that slice.

Why this shape (ticket's recommended option 2)

Keep charging the lookup — it is real work, and charging only files reopens the QA F4 walk-around, where unknown / no-filePath ids cost a free SELECT and move the stall rather than bound it. Instead, size the budget to the true worst case an in-cap agent presents.

Added ARCHIVE_RAW_CURSOR_LOOKUP_MAX (200, mirroring the agent's ARCHIVE_MANIFEST_MAX=200 manifest-entry cap) and set budget = ARCHIVE_RAW_CURSOR_MAX + ARCHIVE_RAW_CURSOR_LOOKUP_MAX. Both terms now mean what they read as; a well-behaved agent (≤200 entries, ≤2000 files → ≤2200 work) is never truncated, while total per-beat work stays bounded at 2200 ops. Corrected the false comment and the warning message; updated .claude/rules/turma-archive.md.

How verified

• Tests: reworked the existing two-term cap test; added an XERK-427 regression proving files spread across transcripts keep every cursor. node --test turma/tests/*.test.js → 1690/1690 pass (archive 41, archive-budget 29, server 700, …). The regression fails against the old single-budget line and passes with the fix (confirmed by reverting only that line).
• QA agent (PASS, no defects): exercised the real module at production defaults (200×10 boundary → all cursors, no truncation warning) and under a hostile 25k-offer manifest (bounded to 2200 ops, ~8ms, 19k dropped with one throttled warning); confirmed the agent-side caps in hub-agent.py bound the producer to ≤200 entries / ≤2000 files; caught 3/3 mutations (old single budget; dropped lookup charge; dropped stat charge); checked positiveEnvInt degenerate inputs (0/-5/abc → 200, can't disable the budget); confirmed no dependents on the old value outside archive.js.