Repository navigation
CodeEraser 1.9.2
CodeEraser 1.9.2
A patch release. Two defects found by an end-to-end verification of 1.9.1
on real repositories are fixed, three report lines are reworded, and one
stale heavy test is refitted. No judgment moves: ce check scores are
comparable with 1.9.0 and 1.9.1, the wire stays at 9.0.0, the daemon
protocol at 2.3.0, the observe feed at 0.13.0, the index schema at 17, the
baseline at ce.baseline/2; no default guard tier moved.
Fixed
- Windows plugin hooks were inert from 1.9.0 to 1.9.1. The plugin
starter takes the PATH core fromcommand -v ce-core; Git Bash answers
that path without.exe. Its SHA-256 matched the pin, so the starter
bound and stamped that string. Spawning it worked (Windows appends
.exe), soce doctorreported a healthy handshake — but the
definition-package loader looked for the literal file, refused, and every
measuring command answeredcore unavailable: no …/ce-core to answer tables/1. The guard, tombstone and audit legs then ran fail-open: on the
project's own feed since the 1.9.0 binding,probe85 degraded / 6
healthy,settle78 / 0,stop_audit30 / 14. 1.8.0 escaped only
because the desktop then ran 1.7.4 and the downloadedce-core.exewas
bound instead.
Fix: the core locator reads a given path the way the spawner does (the
literal file, else on Windows the same path with.exe), and the starter
binds and stamps the file that exists. Covered by a unit leg, an
integration leg (CE_CORE_BINwith and without.exeanswer the same
bytes), and a new bootstrap state that runs a measuring command through
the starter with the realce-coreon PATH. The local stamp
bound-1.9.1.envis not patched by hand; the plugin re-verifies and
rewrites it when it picks 1.9.2 up on its next session. ce precommit/ce commitmsgexited 2 with no reachable core. The
documented behaviour (ce precommit --help,docs/reference/cli.md) is
fail-open: exit 0. Both faces now check the definition package first; when
it cannot be loaded they print the existing fallback sentence, write the
same degraded feed line the hooks write (event=precommit/
commitmsg, no new key), and exit 0. Measuring code is unchanged.
Reworded
ce rules --why --format json: an assertion's proof root printed raw
ids; it now labels its arguments by the goal's sorts (paths), like every
rule row.predsis unchanged.ce erase --applysaid(dry-run; --apply to act)in the summary line
and thenerase applied: …; the summary now says
(applying the eraseable rows)under--apply.ce precommit/ce commitmsgduplicate-block reasons said
ce audit: this session's edits … — deduplicate before stopping; they
now name their own face (the staged changes …) and end
deduplicate before committing. The Stop face is byte-identical.
Tests
moves_line(ignored, ~1 GB of request values) was sized for the retired
32 MiB line and read one body where it asserted two after the line became
128 MiB; it now sizes its fixture from the two ceilings the plan packs by.- The verification itself: the released 1.9.1 binaries over ten seeds, the
project's own clone and sixteen real trees (CLI == MCP, 42 checks, 0
failures); the frozen-oracle replays (language precision ×11, FPR ledgers
×6, zone, L2, flow ×10, tombstone, similar — every tally equal to its
frozen document); sixteen hand-derived items against the methodology
booklets. Old (1.9.1) and new binaries agree byte for byte oncheck,
dedup,deadcode --check,structureanderaseover cobra, requests,
jsoup and luarocks.
Everything else
The sixteen judgments, the four families introduced in 1.9.0 (ce query /
ce rules, ce flow, ce merge, ce arch), the core-side definitions and
report text, and the 128 MiB wire line are as described in the
1.9.0 notes.
The plugin follows the release branch and picks 1.9.2 up on its next
session; Windows users of the plugin should expect the hooks to judge again
from that session on.
Unsigned, as before
The binaries are built by the release workflow and pinned by SHA-256 in
plugin/bin/manifest.env; nothing is code-signed. SHA256SUMS is in the
assets.