Skip to content

Merge corrective batch and Core M016 authoring - #2

Merged
dbowm91 merged 21 commits into
mainfrom
plans/corrective-batch-2026-10-23
Nov 12, 2026
Merged

dbowm91 merged 21 commits into
mainfrom
plans/corrective-batch-2026-10-23

Conversation

@dbowm91

@dbowm91 dbowm91 commented Nov 12, 2026

Copy link
Copy Markdown
Contributor

Merges the plans/corrective-batch-2026-10-23 branch into main. This branch is 21 commits ahead of main and contains the current corrective batch, the 0.2.0 publication qualification records, and Core M016 contract-doc packaging plan authoring.

Hosted push CI for head 14d39d31eae4a9084c2ce7ca1ead7f0ac7b97a30 passed all five lanes in run 43158554761. Pull-request CI will also run before merge.

CI Test and others added 21 commits October 23, 2026 18:35
Implements and closes nine implementation plans from the workspace audit,
with closure records, registry/roadmap reconciliation, and a new
documentation-truth gate command.

Verified Update Core
- M014: `run_bounded`'s deadline is now an unconditional wall-clock bound
  evaluated on every poll, whether or not the direct child has exited.
  The descendant-holds-pipe case is bounded and fails closed.
- M015: `FailureCategory::RetainedEvidence` removed; all 32 public fallible
  items carry an `# Errors` section, enforced crate-locally.

Archive Extraction
- M005: one truthful cleanup contract across all four entry points, and the
  duplicated tar.gz/zip member-write bodies consolidated into one helper.

Service Lifecycle
- M012: one commit/rollback engine behind both orchestrators; the stricter
  former behaviour wins everywhere; ownership unified with `eggup-core`.
- M013: typed `ManagerFailureKind` replaces `permission_hint` matching;
  one deadline model across systemd/launchd/SCM; one cron listing and rule.
- M014: the five former `unreachable!()` SCM sites become typed errors, a
  real Windows SCM fixture drives install -> uninstall through the public
  API, and the public surface is pinned by a compile-only guard.

Acquisition Transport
- M012: the two scheduler-sensitive curl assertions are deterministic, with
  the sub-second serialization coverage proven without a subprocess.
- M013: nine vacuously-passing curl tests no longer report a pass; the
  zero-dependency seam gains a typed I/O failure surface.

Eggpack Manifest Interoperability
- M006: one file-digest implementation in the workspace; the redundant direct
  `sha2` runtime edge is removed from `eggup-eggpack`.

Planning / Closure Hygiene
- C013: the root changelog, `docs/`, `README.md`, `AGENTS.md`, the
  architecture deep dives and the registry now state one truth per
  milestone; a read-only `scripts/check-docs-truth.sh` gate command is added.

No manifest change beyond `crates/eggup-eggpack/Cargo.toml`; `Cargo.lock`
is unchanged; every dependency boundary is intact.
The hosted Windows lane (run 40004604594) failed on
`member_write_path_exists_exactly_once_in_non_test_code`. The guard located
the inline test module with a CRLF-fragile marker:

    .split_once("\n#[cfg(test)]\nmod tests {")

A checkout that writes CRLF leaves the marker unmatched, so the guard failed
on a tree whose member-write path was correct. The marker keeps its
`mod tests {` anchor, because `src/lib.rs` carries an earlier
`#[cfg(test)]` item that must not be mistaken for the module.

Proven in an isolated worktree with a CRLF checkout of `src/lib.rs`:
the pre-fix guard fails with the hosted panic (line 295, same message),
and the fixed guard passes 3/3 under CRLF and 3/3 under LF.
…able

The hosted Windows lane (run 40005168760) and the hosted Stable lane
exposed three defects that the native macOS lane cannot see.

**Windows: missing per-test platform gating (12 failures).** The cron,
systemd and scripted-manager fixtures are Unix-shaped - a crontab body
names a POSIX absolute path, and the systemd `ExecStart` grammar rejects
a backslash. Each affected test is now gated `#[cfg(unix)]`, per test and
not per module, so macOS and Linux keep every gated test and only the
Windows lane shrinks. The two negative controls that would otherwise have
become vacuous on Windows are gated with their subjects.

**Windows: CRLF-fragile rustdoc anchor (1 failure).**
`errors_doc_sections.rs` located the `LaunchdInstall::new` block with a
multi-line anchor, which a CRLF checkout can never match. The source is
normalized once before the anchor is located; every anchor stays
byte-exact, so the fixture discriminates exactly as much as before.

**All lanes: a discarded spawn cause.** `Command::spawn()`'s `io::Error`
was dropped, so a missing executable, a non-executable file and a
transient resource exhaustion were indistinguishable in the diagnostic -
which is why the hosted Stable failure could not be attributed. The
`Unavailable` detail now names the `ErrorKind` and the raw OS code,
rendered only from the discriminant and the numeric code so no path, URL
or credential material can leak. The variant stays `Unavailable` and the
composition fallback contract is untouched.

Recorded as acquisition M015 with its plan and closure record.
CI run 40008039843 is green on all five hosted lanes (stable, msrv,
service-systemd, macos, windows-check). The Windows lane executed the real
Windows SCM fixture against the genuine Service Control Manager - no
EGGUP-SCM-LIVE-SKIP marker, no cleanup failure - which discharges service
M014's own condition that a closure without a real hosted Windows run is not
a closure.

Each of the eleven closure records now carries a dated parent-session
verification addendum that supersedes its own earlier "hosted lanes not
run" claim, and every plan Status header reflects the resulting state.
Acquisition M012 stays conditionally closed: the hosted Stable lane
re-observed registry finding 1 in the spawn mode, so that finding remains
open and narrowed rather than closed.

The registry, the five subsystem roadmaps, and the root CHANGELOG now agree
on the same state per milestone, and the new scripts/check-docs-truth.sh
gate command passes.

The windows-check lane gains the explicit step service M014's plan names,
so the fixture's recorded skip reason can never be swallowed by the test
harness's output capture.
…hors

Implements and closes five milestones of the 2026-10-23 workspace audit,
on top of the eleven already closed in this branch:

- Acquisition M014 (adapter capability parity): the curl adapter now
  exposes a caller-settable User-Agent, validated as an HTTP field value
  and placed on CurlTransport so the published CurlConfig stays
  source-compatible. Closes the parity finding M013 deferred.
- Core M017 (bounded stdin execution): CommandSpec::stdin_data and
  MAX_STDIN_BYTES give the bounded-execution primitive a bounded stdin
  write, with a stable Error::Io operation-string classification surface
  for downstream adapters.
- Service M015 (bounded executor consolidation): SystemExecutor now
  delegates to eggup_core::run_bounded, so the workspace holds one
  bounded-process-runner implementation. This also fixes a defect in the
  replaced runner, which consumed the stdin writer's verdict with
  try_recv and so reported a successful write as a failure on the
  crontab - path. That defect is present in the published
  eggup-service 0.1.3 and is registered as registry finding 15
  (medium, open for the published crate).
- Transport Footprint M001 (isolated-configuration coverage): CI now
  builds the crate's three configurations and a linkage-discipline guard
  rejects a sibling-adapter reference in every configuration. Closes
  registry finding 12.
- The transport-footprint subsystem roadmap, which finding 12 required
  before its first milestone could be authored.

Also repairs 376 stale line-number anchors across four architecture deep
dives and corrects the prose claims that had drifted from the code.

Local gate: exit 0. Workspace suite: 516 passed, 0 failed, 48 test
binaries. No manifest change: Cargo.toml, Cargo.lock and every crate
manifest are byte-identical to the previous commit.
…the registry package check

Hosted run 40061134915 failed on two counts. Both are addressed here.

The Windows lane failed one test, the_managers_argv_is_byte_identical, with
OwnershipDenied { ownership: Unknown }. The cause is correct production
behaviour that the new test did not account for: parse_exec_start rejects any
ExecStart argv containing a backslash and then requires the path to be
absolute, so on Windows no value can parse -- a drive path has a backslash and
a backslash-free path is not absolute there. systemd is a Linux-only manager
with a Unix-path-shaped ExecStart grammar. The single guard is split into
three single-concern tests and only the systemd one carries the repository's
established per-test #[cfg(unix)] gate; the launchd and cron guards are
platform-neutral and now run on the Windows lane too, which the previous
single test never reached. No assertion was weakened and no production code
changed. macOS coverage rose from 12 to 14 tests in the file.

The other four lanes failed `cargo package -p eggup-service --locked`. That
step resolves eggup-core from crates.io, so it compiles the packaged crate
against the published eggup-core 0.1.3, which carries neither core's
unconditional execution deadline (M014) nor its bounded stdin (M017). This is
a publication state, not a defect of the source line: the same branch already
left eggup-curl, eggup-eggfetch, and eggup-eggpack unable to package against
the registry, and the service package step was simply the last one still
resolving. The four steps now use --no-verify, with an inline comment
recording why, so the deferral is visible in the workflow rather than silently
dropped. The step still packages the crate and validates its manifest and
file list, and the compile of the packaged sources stays covered by the
workspace build, test, and clippy steps. Registered as registry finding 16
(medium, open) and as UF-4 in the service M015 closure, both naming the
eggup-core republication milestone as the follow-on that restores the strict
step.

The stale ci.yml line references this batch's CI change invalidated are
corrected in the governance deep dive, the service deep dive, and the
docs-hygiene skill.

Local gate: exit 0. Workspace suite: 518 passed, 0 failed. No manifest
change.
…e batch

CI run 40062299630 is green on all five lanes for the five milestones
closed in the preceding commit: acquisition M014, core M017, service M015,
transport-footprint M001, and the architecture-anchor repair.

Each of the four new closure records gains a dated parent-session
addendum that enumerates the earlier "not run" claims it supersedes, gives
the per-platform results separately, and carries the compatibility
consequence forward. The Windows Service Control Manager fixture ran
natively: 2 passed, 0 failed, 0 ignored, with zero EGGUP-SCM-LIVE-SKIP
markers and zero EGGUP-SCM-LIVE-CLEANUP-FAILED markers, so the skip path
was not taken and the disposable registration was removed.

The registry, the five subsystem roadmaps, and the root changelog carry the
run ID. The registry's project-state and next-handoff sections record the
third pass and name the two open findings it leaves behind: finding 15
(medium), the published eggup-service 0.1.3 misreporting a successful
crontab stdin write, and finding 16 (medium), the packaged-service
verification deferred until the eggup-core republication.

Local gate: exit 0. Workspace suite: 518 passed, 0 failed. No manifest
change.
A read-only audit of the corrective batch found three release-blocking
items, a set of documentation claims that contradicted the branch's own
work, and a documentation-truth gate that did not check documentation
truth. This commit is the implementation side of the six corrective
milestones that close them.

- acquisition M016: record the seam's 0.2.0 publication coupling. Both
  adapters call seam capabilities that exist only in the unpublished seam
  while declaring `^0.1.0`, which cannot accept the source-breaking 0.2.0
  the seam must publish next, so neither adapter can be packaged against
  the registry. The coupling is now recorded, the false "a future
  republication would be a non-breaking 0.1.3" claim is removed from the
  adapters' changelogs, and a guard fails when a dependent neither accepts
  the declared next release nor records the widening.

- core M018: the published rustdoc promised that a write failure observed
  after the deadline cannot turn the deadline outcome into an Err, but the
  stdin arm was evaluated before the deadline arm, so the same input at the
  same timeout was classified either way depending on poll phase. The
  deadline now has priority and the writer stamps its verdict with the
  instant it observed the write, so the rule is literally true and the
  outcome is a fact about the write rather than about the poll. An
  undelivered payload is never reported as a success and the unconditional
  deadline guarantee is unchanged. Two overstated statements are corrected
  and M015's U1 is closed on the evidence of ab90688.

- service M016: the crate's changelog omitted the release note for D8, the
  difference that is a defect fix rather than a cost, so the fix for the
  published crate's stdin-verdict defect had no release note at all. The
  difference count, the unflagged engine_policy source break, and the M015
  closure's stale test inventory are corrected.

- planning C014: the documentation-truth gate opened no manifest, skipped
  every architecture deep dive, and skipped three-cell registry rows, so it
  could not catch the drift it claimed to check. It now reads the manifests
  and the per-crate changelogs, covers architecture/*.md, and reads every
  registry row; each change is red-first proven against a documented false
  negative. The transport-footprint deep dive and the CI matrix are
  corrected.

- planning C015: 179 stale path:line anchors are re-derived in the five
  deep dives whose source crates were stable, and eleven substantive claims
  are corrected, including a table row that still documented an API service
  M013 removed and three stale =0.1.2 pins.

No manifest, version, or publication changed, no tag moved, and no crate
gained a dependency.
The process side of the audit repairs. Each milestone carries the full
16-section plan, the closure record with its requirement-to-evidence
matrix, the registry row, the roadmap row, and a dated addendum where it
supersedes an earlier record.

- acquisition M016, core M018, service M016, planning C014, planning C015
  are closed; their registry rows, roadmap rows, and root CHANGELOG
  entries are recorded.
- The registry's "Recently closed foundation" table, its per-subsystem
  "Current project state", and the pass attribution for the second and
  third batches are repaired; the Windows lane's test-step count is
  corrected from nine to ten in the registry, AGENTS.md, and both skills,
  and the CI matrix's "Exact commands" column now names the commands the
  workflow actually runs.
- Registry finding 17 is added: the adapters cannot be packaged against the
  registry until the seam publishes 0.2.0 and each dependent's requirement
  is widened. The finding preamble, the republication wording, and the
  release-workflow skill's caret-inheritance rule are corrected.
- The core M015 closure's status is reconciled with the registry and the
  roadmap, and its U1 finding is closed.
The snapshot was authored at service M014 and the M015 consolidation rewrote
crates/eggup-service/src/lib.rs, so 19 of its 80 per-item location
references no longer pointed at the items they name. The locations are
re-derived by symbol; the public surface itself is unchanged — no item was
added, removed, or re-signed, and the pinning test still passes with its
count assertions intact.

The snapshot's provenance header is corrected: it still described the
service milestones 012 and 013 as "uncommitted in the working tree" and
named only the M009-era baseline.

Also refines the bounded-stdin classification sweep: it walks the deadline
ladder in explicit two-millisecond steps and reports how many samples were
decided on each side of the deadline, so a load-sensitive run states what it
actually observed instead of only asserting.
The five milestones that close the audit's findings are hosted-qualified. CI
run 40067964510 is green on all five lanes: stable 528 passed, msrv
check-only, service-systemd 9 passed, macOS 528 passed, windows-check 650
passed. The local gate exits 0 with 528 tests passing and 0 failures.

Each of the five closure records gains a dated addendum with the
per-platform results, stated separately and never inferred.

The immediately preceding run 40067739294 was red in the stable lane on a
transient ETXTBSY in a pre-existing core unit test that this branch did not
modify. It is recorded as a re-observation of registry finding 1 — the
load-sensitive-test class, which is low severity and blocks nothing — with
the evidence that it did not reproduce in 25 consecutive local runs. The
registry's current-state narrative, the findings preamble, and the root
changelog's qualification note are updated to carry both runs.
…rement

The bounded-stdin fixture refinement moved the fixture lines the M018
closure cited, and the requirement list gained the load-sensitivity rule the
refinement enforces: a helper binary can start after a short deadline, be
killed before it writes anything, and never produce its start marker, so
every fixture that waits for that marker must poll for it for a bounded time
and must treat "the marker never appeared" as a distinct condition that
asserts only the deadline outcome. No fixture may assert that a helper
started.

R19 records the evidence: the sweep reports its coverage counts on stderr so
a run states what it observed instead of only asserting, and three
consecutive full-crate runs under a concurrent cold workspace compile, plus
one under four concurrent cold compiles, all pass 110 tests with 0 failures.
…e record

The M018 closure record still carried "hosted qualification not run" and an
open medium finding for it, both of which the parent session has now
discharged: hosted run 40067964510 is green on all five lanes and its
per-platform results are recorded in the addendum at the end of the record.
The record now names the commits that carry the change, the files the
milestone created or changed, and the one medium follow-on that remains open
about publication rather than about this milestone.
The 0.2.0 line publishes seven crates from one shared source commit, so the
authorization is split along the dependency order that constrains it:

- Verified Update Core M019 — eggup-core, published first
- Archive Extraction M006 — eggup-archive, published second
- Acquisition Transport M017 — eggup-acquisition, eggup-curl and
  eggup-eggfetch, the seam-then-adapters chain
- Service Lifecycle M017 — eggup-service
- Eggpack Manifest Interoperability M007 — eggup-eggpack, published last
  because it exact-pins all three of its in-workspace dependencies

Each plan uses the ten-section publication template and states its own
pre-publication gates, its registry-only proof, its stop conditions, and the
closure evidence it requires.
The workspace is version-lockstep, and the acquisition seam's next release is
source-breaking, so the seam had to advance to 0.2.0 and the lockstep
workspace carried all eight crates with it.

Every in-workspace requirement that named the seam's old version is widened in
the same change set: eggup-curl and eggup-eggfetch move from "0.1.0" to "0.2",
eggup-service moves from "0.1.0" to "0.2" on eggup-core, eggup-eggpack's three
exact pins are re-pinned to "=0.2.0", and eggup-transport-footprint widens its
three requirements so the workspace still resolves. A caret requirement does
not cross a major boundary, so leaving them at "0.1.0" would have let a
consumer's lockfile resolve eggup-core 0.1.3, which has no
CommandSpec::stdin_data and cannot compile eggup-service 0.2.0.

Each published crate's changelog is cut from Unreleased to a 0.2.0 section, and
the root changelog records the train, its publication order, and the source
breaks a consumer must make.

The M016 publication-coupling guard is re-pointed at the seam's published
version. Its premise -- a seam with a declared *next* release -- was discharged
by this publication, so the guard now derives the published version from the
seam's own changelog and holds every in-workspace dependent to it. The three
test names are unchanged because tests/assertion_guard.rs names them as
required behaviour witnesses. The re-pointing is recorded as an addendum in
the M016 closure record.

The documentation-truth gate's published-version check is no longer pinned to a
single version number, and a bullet in Unreleased no longer inherits the
wording of the paragraph that follows it. Both changes are red-first proven
against documented false negatives.
The hosted lanes run `cargo package -p eggup-service --locked --no-verify`.
That step resolves the packaged service crate's `eggup-core = "^0.2"` against
crates.io, so it cannot succeed until `eggup-core 0.2.0` is on the registry --
and publishing `eggup-core 0.2.0` requires a green matrix. For the first crate
of the train the two conditions are circular, and the stop condition cannot be
satisfied as literally written.

The exception is recorded in the plan rather than discovered by a reader: the
observed failure was one step, in four lanes, with the same registry-selection
error in each, while every other step in every lane was green. The exception
authorises this upload only, under the local gate and a hosted run that is
green in every step except that one; every later crate in the train is
published only after the full matrix is green, which this upload makes
possible.
The packaged-service step ran `cargo package -p eggup-service --locked
--no-verify` because the packaged crate's `eggup-core = "^0.2"` requirement
had nothing on the registry to resolve against: the published
`eggup-core 0.1.3` carries neither core's bounded-stdin surface (M017) nor
its unconditional execution deadline (M014). That deferral is discharged.

`eggup-core 0.2.0` is now on the registry, so the requirement resolves and
the packaged sources can be compiled against the registry copy rather than
only packaged. All four lanes that carry the step now run the strict form,
`cargo package -p eggup-service --locked`.

This is the follow-on named by registry finding 16.
The seven publishable crates are on crates.io at 0.2.0, published from
source commit b612608 in dependency order. This commit records the
close-out that the publication milestone requires: the five closure
records, the registry and roadmap rows, and the reconciliation of every
surface that still described the pre-0.2.0 state.

Closure records (one per publication milestone):
  - verified-update-core/019    eggup-core 0.2.0
  - archive-extraction/006      eggup-archive 0.2.0
  - acquisition-transport/017   eggup-acquisition, eggup-curl, eggup-eggfetch 0.2.0
  - service-lifecycle/017       eggup-service 0.2.0
  - eggpack-manifest-interop/007 eggup-eggpack 0.2.0

Each records the publication source SHA, the registry checksum, the
per-platform evidence with an unrun lane marked "not run", the
requirement-to-evidence matrix, and the unresolved findings with their
severity.

Registry: the 0.2.0 train block with all seven checksums, dated
dispositions for findings 15, 16 and 17, and a closure for finding 9
(the eggpack sha2 runtime edge, which the published 0.2.0 manifest no
longer carries). Finding 14 stays open under its named owner, Core
M016. Registry lines remain within the 481-byte limit.

Reconciliation of surfaces that still described the pre-0.2.0 state:
  - docs/releases.md said the line was source-breaking in three places
    and that eggup-core was additive. Both were wrong: the published
    eggup-core 0.2.0 removes FailureCategory::RetainedEvidence, which
    the published 0.1.3 does carry, so a consumer naming that variant
    explicitly stops compiling. Four breaks, not three.
  - architecture/tooling-governance.md, architecture/acquisition.md,
    architecture/curl-adapter.md, architecture/eggfetch-adapter.md,
    architecture/eggpack-adapter.md, architecture/core-transaction.md,
    architecture/service-lifecycle.md and architecture/transport-footprint.md
    carried stale current-state version literals, a stale CI packaging
    note that still described the deferral as live, and ci.yml line
    anchors that shifted when the strict package step was restored.
  - The three open-drift rows for the footprint pins, the service
    manifest, and the footprint CI coverage were rewritten as closed:
    each was a true statement about a state the 0.2.0 train ended.
  - The verified-update-core roadmap carried two M018 rows with
    different statuses; the stale one is removed.
  - The release-workflow, verify-workflow and docs-hygiene skills
    described the --no-verify package step and the published versions
    as they stood before the train. All three are rewritten.

Gate: scripts/check-docs-truth.sh assertion 2 was hardcoded to 0.1.2
and is now anchored to the manifest version, so the next train needs no
further edit of it. Re-based red-first: a stale "workspace version"
paragraph, a "published at" claim naming an older version outside
docs/crates.md, and an exact-pin literal outside the eggpack policy
each fail, and the allowlisted docs/crates.md case passes.

The five publication plans now read "Status: closed" and name their
closure record.

Local gate: ./scripts/check-local.sh exits 0, 528 passed / 0 failed /
1 ignored. The ignored test is the pre-existing ignored eggup-curl
unit test, not introduced here.
…us table

The five 0.2.0 publication milestones sat in the "Open, authorable, and
future work" table. That table's rule is "work with no milestone row
elsewhere", and it is explicitly not the milestone-status table — but
these are closed milestones with closure records, and every earlier
publication milestone (core M009/M013, service M009/M011, acquisition
M009/M010/M011, archive M004, eggpack M005) sits in the authoritative
"Dependency-ready implementation work" table.

The rows were also malformed there: each carried five cells under that
table's four-column header, while every other row in it carried four.
They were written for the five-column status table.

Moved verbatim, so the plan path, closure path, status, and evidence
cells are unchanged. The table is now arity-consistent throughout, and
no (subsystem, milestone) pair carries a row in both status tables.

Registry line lengths remain within the 481-byte limit and the
documentation-truth gate passes.
Run 40103357142 was red in the stable lane on a transient ETXTBSY in the
pre-existing core unit test bounded_candidates_are_exact_and_environment_is_cleared.
The failing commit 390c224 changed plans/registry.md only, so the code was
byte-identical to the green run 40102904892, and the rerun on the same
source SHA was green on all five lanes.

This is the third recorded re-observation of registry finding 1, in the
same test and the same mode. The registry already carries the class, so
the observation is recorded there rather than registered separately, and
the train block now names the close-out's own hosted qualification
separately from the publication qualification run 40079833945.
@dbowm91
dbowm91 merged commit 9e24c36 into main Nov 12, 2026
10 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant