fix(context): the audit's control and fit, and rung 1's price

[matiasmacera]

What

Three defects in the two tools that measure context loading. /aios:housekeeping runs both.

#	Where	What happened
1	hooks/context-load-audit.py	Primary sessions are the audit's control. They were recognised by a regex of session names from a single operator. On any other vault those names match nothing, so the real primary is counted as a worker and can be listed as a suspect.
2	hooks/context-load-audit.py	The fit check flags a worker that acted outward without reading a declared/ file. It only scanned the first --cap (120) calls, so a later export was never seen. It also ignored order: a worker that exported first and read the voice file afterwards counted as having read it.
3	hooks/context-rungs.py	Rung 1 is defined as what context-floor.py emits. The floor emits a rule library's index and the newest entries of each observed file by date. Rungs priced the last five entries of every file, rule library included. It defined a rule-library regex it never used.

Fix

1. Primaries come from the first table under ## Identity in USER.md, the table that already decides how a session is greeted, plus --primary NAME. The parser follows GFM (a header row, then a delimiter row). It skips the template's italic example rows and ignores fenced code, and the table ends at the first line that is not a row. With no primary declared, the audit aborts and says how to declare one.
2. Fit reads the whole transcript. It flags a worker whose first outward action comes before any declared/ read. When one call does both, the positions of the matches that produced the hit decide. --cap and --min-tools still apply to the loading window; the JSON gains tools_total and declared_before_outward.
3. The floor's per-file selection moved into one function, recent_slice(). context-rungs.py imports it, and its recency constant, from the floor, so price and emission cannot drift. The floor's output is byte-identical before and after, in text and --json, with default, zero and widened --recent.

Proof

• tests/context-rungs.test.sh builds a vault with a rule library and a chronology whose tail is its oldest part. It asserts rung 1's per-file price equals the words the floor emits for each file. Against the pre-change hook, pinned by sha, the prices differ by two orders of magnitude.
• tests/context-load-audit.test.sh covers declared primaries, the no-USER.md abort, --primary, and 20 parser cases. It also covers fit past the cap, a read after the ship, a ship then read inside one call, and a look-alike filename earlier in the same call. --min-tools still filters on the loading window. The pre-change hook, pinned by sha, first has to produce a valid report, and then it flags none of the fit cases.
• Every check has a mutation control: reverting it fails exactly its case.
• tests/spawned-worker-context-parity.test.sh compared two numeric recency constants. It now also accepts rungs reading the floor's constant, which is the stronger form.
• The full tests/ run passes. The three suites also pass under bash 3.2.

Scope

hooks/context-load-audit.py, hooks/context-rungs.py, hooks/context-floor.py (refactor only), the three tests, hooks/_index.md, the Bucket 30 text in housekeeping.md, and the CHANGELOG entry. It continues the hooks series announced in #158.

🤖 Generated with Claude Code

[chuycepeda]

Shipped in v0.8.5 with your commits and authorship intact. Thank you. It went in as you wrote it. The floor refactor was checked against a live vault: context-floor.py output was byte-identical before and after, across six argument sets (107–297 KB). On that vault the audit now finds exactly one primary, where the old pattern also counted four workers as primaries.