the reader retries a race, and the last in-place producer stops making one by richardkiene · Pull Request #87 · Liquescent-Development/gascan

docs: implementation plan for the reader's retryable verdict

richardkiene

committed

3e2cc9e

refactor: the staging prefixes are shared protocol, because there are two stagers now

richardkiene

committed

6c42e12

fix: the staging sweeper covers both stagers, and says why the lock is what makes it safe

richardkiene

committed

409f1e1

docs: the sweeper's lock argument names all three locked callers and its own residue

richardkiene

committed

0b473fb

fix: retirement stages and renames rather than mutating a live record in place

richardkiene

committed

6d040c2

fix: retirement proves the name still holds the record before it replaces it

richardkiene

committed

22f4bd4

test: a concurrent observer proves the reclaim path shows only legal faces

richardkiene

committed

9b8cf22

feat: race-shaped reader failures carry a type, and everything else stays terminal

richardkiene

committed

d25d218

fix: the reader retries a race instead of reporting it as a fault

richardkiene

committed

5f2b800

test: the race marker's journey out of an observation is pinned end to end

richardkiene

committed

7f592ca

fix: a tombstone unlinked by a successor is a race rather than a permission fault

richardkiene

committed

10bf20f

test: a symlink over the tombstone stays a fault, and the ENOENT split says what it changes

richardkiene

committed

beb05f4

docs: the durable text says what is true now, and what was traded to get there

richardkiene

committed

4931a0d

docs: a count that belonged to the file, and three anchors that had drifted

richardkiene

committed

24cdf53

docs: the two literals sit in DaemonLaunch, not DaemonPaths

richardkiene

committed

4b72c43

docs: the reader's residue is five failures, and the common case is not among them

richardkiene

committed

dbdaf1d

test: the retry and the observation are pinned together, not only apart

richardkiene

committed

409fa9e

docs: a fourth CI flake mechanism, and it is the pid file read while empty

richardkiene

committed

a902fd5

fix: a swept reclaim staging file names itself instead of a bare ENOENT

richardkiene

committed

1a6611a

fix: readiness polls a never-settled inspection instead of failing at once

richardkiene

committed

97f8a4c

docs: a fifth CI flake mechanism, and this one is main's own

richardkiene

committed

56b343c

docs: the review's five minors reach the tree, and one of them is cheaper than the file implied

richardkiene

committed

d5ea334

docs: PR #87 is open by choice, and finishing the reader half is the assignment

richardkiene

committed

b1ef129

docs: a sixth CI flake mechanism, and this one failed on a commit with no code in it

richardkiene

committed

bf107a1

fix: the reader classifies every window a stop transition tears across

richardkiene

committed

ae03597

docs: the reader half is complete, and the residue list it left was itself incomplete

richardkiene

committed

8613f22

fix: the reader's completeness claim was wrong, and a review found it on the production path

richardkiene

committed

93c77fe

docs: a seventh CI flake mechanism, and it is a second distinct one in the same target

richardkiene

committed

cfcfb62

Navigation Menu

Uh oh!

the reader retries a race, and the last in-place producer stops making one - #87

the reader retries a race, and the last in-place producer stops making one#87
richardkiene merged 34 commits into
mainfrom
fix/daemon-reader-retryable-verdict

Commits on Aug 19, 2026

docs: implementation plan for the reader's retryable verdict

refactor: the staging prefixes are shared protocol, because there are two stagers now

fix: the staging sweeper covers both stagers, and says why the lock is what makes it safe

docs: the sweeper's lock argument names all three locked callers and its own residue

fix: retirement stages and renames rather than mutating a live record in place

fix: retirement proves the name still holds the record before it replaces it

test: a concurrent observer proves the reclaim path shows only legal faces

feat: race-shaped reader failures carry a type, and everything else stays terminal

fix: the reader retries a race instead of reporting it as a fault

test: the race marker's journey out of an observation is pinned end to end

fix: a tombstone unlinked by a successor is a race rather than a permission fault

test: a symlink over the tombstone stays a fault, and the ENOENT split says what it changes

docs: the durable text says what is true now, and what was traded to get there

docs: a count that belonged to the file, and three anchors that had drifted

docs: the two literals sit in DaemonLaunch, not DaemonPaths

docs: the reader's residue is five failures, and the common case is not among them

test: the retry and the observation are pinned together, not only apart

docs: a fourth CI flake mechanism, and it is the pid file read while empty

fix: a swept reclaim staging file names itself instead of a bare ENOENT

fix: readiness polls a never-settled inspection instead of failing at once

docs: a fifth CI flake mechanism, and this one is main's own

docs: the review's five minors reach the tree, and one of them is cheaper than the file implied

docs: PR #87 is open by choice, and finishing the reader half is the assignment

docs: a sixth CI flake mechanism, and this one failed on a commit with no code in it

fix: the reader classifies every window a stop transition tears across

docs: the reader half is complete, and the residue list it left was itself incomplete

fix: the reader's completeness claim was wrong, and a review found it on the production path

docs: a seventh CI flake mechanism, and it is a second distinct one in the same target

Commits on Aug 20, 2026

docs: the fifth mechanism recurred, and again on a commit with no code in it

docs: an eighth CI flake mechanism, and this one is a wall-clock bound rather than a race

fix: the second review's findings, including two claims the last commit made and did not keep

docs: hand off with the reader half closed, both reviews discharged and CI green

docs: commit both review rounds and correct two counts the handoff left stale

docs: refresh the cold-start SHAs and record that PR #87 is MERGEABLE/CLEAN