mirror of
https://github.com/stablyai/orca.git
synced 2026-09-23 00:02:29 +00:00
Jinwoo Hong
62903b7cf6
* test(mobile): pin the C5 page closure golden by golden (OTA phase C, C5.2)
The suite next door proves the corpus replays byte-identically or in a named
class, as counts over 787 goldens. A count cannot tell this domain's regression
from another domain's improvement, so the 134 recorded at a call site the
agent-history page owns are pinned by id, the way C1's 103 are.
Re-derived at this base rather than taken from the design: 3510 modules, 370
local, 16 under `src/agent-history`, 27 families, 134 goldens. The design measured
3508 / 368 at `889c2b562f`; the two new local modules are `omp-draft-launch.ts`
and `omp-fresh-launch.ts`, which arrived on main, and no family or golden moved.
The `.web.tsx` sibling is the entry: the native file reaches `MobileWebShellScreen`
and pulls the shell in, which adds 46 local modules and the two `mobileWeb.*`
families the C1 docstring excludes by name — so measuring the wrong one of the two
gives the right pin table and a wrong module count.
C1's 22 families are inherited verbatim rather than re-derived. Deriving them from
C2's classification rule disagrees with the committed file on 10 of its 103 pins,
in three ways that are each a true statement read too far: `tasks.smart-source
-search` is a `params-undefined` family and the rule's list of those is the five
C2 adds, not C1's one; `host-worktree-refresh` carries `write-ordinal` and
`result-absent-stream-release` goldens, classes the rule does not model because no
family C2 adds carries one; and "a scenario scripting `{ ok: true }` with no
result" is a property of the scenario a golden derives from, not of its family, so
reading it family-wide moves `worktree-catalog-snapshot`. The rule decides the
five families C5 adds and nothing else.
`c5PageClosureTotals` is pinned beside the per-id walk, and the gate asserts the
run's own class counts over this closure against it. A per-id walk agrees with a
table that is wrong the same way twice: deriving this file with a reader that
skipped `c1-page-closure.ts`'s wrapped entries produced an empty mismatch list, 53
`result-absent-settlement` and 0 `result-absent-stream-release` where the truth is
50 and 3. Only the counts showed it, so only the counts can guard it.
Both halves proved red, not assumed. A verdict changed in the pin fails the gate
as "aivault-history-scan-fulfilled: pinned params-undefined, ran identical"; a
golden dropped from the pin fails three ways — the drift names it, the totals move
72 to 71, and the census count fails.
Tallies untouched: 787 goldens, 396 / 341 / 3 / 6 / 33 / 8, unclassified 0. No
golden moved and nothing under `mobile/rpc-foundation/` was written.
Claude-Session: https://claude.ai/code/session_01JNnE9qzUZMMnqpZWCqM3nb
* refactor(mobile): give both page closures one drift, not two copies
Round 1 on #21649. C5's pin file arrived with C1's type alias, observation type,
exclusions walk and drift walk copied across, identical but for the names. The
copies are the problem rather than the duplication: these pins are the instrument
two domains argue with when a golden moves between families they share, and C5's
closure contains C1's 22 families entire — so two spellings of "what drifted" can
disagree about the same goldens while both stay green.
`page-closure.ts` now holds the types and four helpers, each taking the pins as an
argument; `pageClosureTotals` moves there too, since it is a general question and
was C5's only real addition. `readPageClosure` joins them: the gate had grown a
local helper to print one domain's line, which is the same shape one step later.
Both pin files are now their table and nothing else, which is the part that is
per series and meant to be read as a diff.
Output asserted rather than assumed. The gate's two closure lines are byte-
identical before and after — "C1 page closure: 103 goldens in 22 families, 54
byte-identical" and "C5 page closure: 134 goldens in 27 families, 72
byte-identical" — diffed as files, and both pin suites report the same 60 cases
with the same verdicts.
Claude-Session: https://claude.ai/code/session_01JNnE9qzUZMMnqpZWCqM3nb
* test(mobile): read a golden's family through the corpus's own reader
Round 1 on #21649. Both pin tests parsed a golden file and asserted its shape in
order to read one field, each with a `SAFETY` annotation saying every golden
carries `family`. `readGolden` already answers that question properly: it checks
the format version and the value pool and throws a named diagnostic otherwise,
and returns `family` typed.
Both annotations are gone rather than reworded. C1's is fixed in the same commit
because it is the same line — the cast arrived in C5's file by being copied from
it, and leaving the original would have kept the shape that produced the copy.
The assertion is also stronger than it was. Before, a golden whose format version
had moved would still be read and its `family` compared; now that file fails with
the reader's own message instead of being silently accepted.
Checked that it still discriminates rather than assuming: renaming one family in
the pin fails two cases, the corpus-family read among them.
Claude-Session: https://claude.ai/code/session_01JNnE9qzUZMMnqpZWCqM3nb