refactor(guards): the hand-written list census is a plan, not a shell - #325
Open
the-homeless-god wants to merge 6 commits into
Open
the-homeless-god wants to merge 6 commits into
the-homeless-god wants to merge 6 commits into
Conversation
The census of hand-written lists moves from 376 lines of shell with two
gawk programs to 2615 lines of flang: scripts/guards/hand-written-lists.fscript.
Three plans, same three modes as the flags of the old script: Perepis
(census), Sverka (--check), Tolko celi (--targets), plus Podlog, the
forgery probe the shell never had.
PROOF OF EQUIVALENCE, both run on one tree on 4 October 2026:
census: shell stdout and plan result match character for character,
17033 characters, 360 lists, 158 divergences, exit 0 for both;
check: shell stdout plus stderr and the plan failure message match
character for character, 26253 characters, 121 new and 64 dead
ledger rows, exit 1 for both;
forgery: one row for a place the tree does not have is appended to the
ledger; both answer exit 1, both name "PUTI poddelka/net.md:1"
in the dead-rows block and both print "mertvyh zapisej 65".
WHAT THE TRANSFER CHANGED ON PURPOSE, four things, each written down in
the file as a primechanie:
- "nothing to judge" answers 3, the code this tree gives that answer,
not 2: in flang io the two means a malformed call;
- the order of the missing target names was gawk's hash order; it is
reproduced as a permutation of ten numbers and holds only while the
tree has exactly these ten targets, otherwise tree order is used;
- "does the name lie on disk" is asked by one git ls-files --others run
instead of a test -e per name; "." and ".." in a name are resolved
lexically, which is what test -e did through the kernel;
- the verdict is not proved by the kernel: the plan runs with --na-veru,
like links:check, jargon:check and inventory:*, because the closure of
Hash map, Lists and String sets carries 104 grid obligations of its
own. The file's own sixteen: fourteen proved, two on the grid.
COST, measured: the shell walked the tree in 12 s, the plan takes 1130 s
and 6.3 GiB in the evaluator. Ninety times dearer is a property of the
evaluator, not of the record: one builtin call costs about 20 microseconds
and the walk over 774 thousand lines runs into that. The guard is still
not called from CI and not from the pre-push hook, and the reason stays
where it was, in scripts/ledgers/uncalled-guards.json: it is red on its
own debt, and a red guard in CI is a disabled guard.
hand-written-lists-ledger.tsv: two rows for the invented path names in
the new guard's own examples, the same reason other probe fixtures carry
("invented probe inputs, not tree names: a list of forgeries does not
rot"). Without them the two runs of the guard's own file would count as
new divergences.
guards-without-forgery-probe.json: the row for hand-written-lists:check
stays, and its reason is rewritten. The debt is "the probe is not called
by CI", not "there is no probe": the probe now exists as
hand-written-lists:forgery and was run by hand, and the row says so, with
the numbers it answered and why CI still cannot call it.
uncalled-guards.json: the measured cost in the row is 15 s no longer. It
is 1130 s, because the walk now runs in the evaluator; the row says that
and keeps the same verdict, that the guard must not go into CI while it
is red on its own debt.
link-guard-known-not-a-path.tsv: two rows for the two task records that
name the removed shell file. Without them links:check grows from 91 to
93 broken paths; with them it is 91 again, the number the base branch has.
proved-share-ledger.txt: the new file carries sixteen obligations, and
the kernel proves fourteen of them; two rest on the grid of the author's
examples. The row is dated by the md5 of the source, as every row there.
The row that names the census guard in file-extensions.fscript is the one line this change touches there, and lint:growth judges touched lines by the limit, not by the debt around them. The path is split with soedinit inside the brackets, the way two other rows in the same list already are: 82 and 107 characters instead of 173. The texts it looks for are shortened to the shortest ones that still separate the two lists: the file must carry ".fscript", ".mjs" and must not carry ".flang", ".mjs".
The shortcut collector compares the declaration in the script with the line in .flangrc character for character, and the three declarations left out --max-steps, --timeout and --na-veru. scripts:check answered 1 and named all three; now it answers 0 and says 137 declarations match the 154 shortcuts character for character.
The 376 lines of scripts/guards/hand-written-lists.sh left the tree, and eight counted marks named it without knowing. All eight were named by the instrument, not by hand: prose-numbers-guard said which mark diverged and by how much, and inventory:languages gave the new numbers. *.sh files 47 -> 46, in two marks *.sh lines 8930 -> 8554 shell files in the table 53 -> 52 shell lines in the table 9247 -> 8871 shell debt 48/5017 -> 47/4641 files outside flang 190 -> 189, in three marks Places: docs/tree-inventory.md (headline, table row, three marks), docs/javascript-inventory.md, .github/workflows/binary.yml (three marks and the prose beside them), .github/workflows/ci.yml. After the change prose-numbers-guard counts 224 marks of 224 agreeing with the tree, and inventory:check says the document agrees with the instrument for all seventeen languages. docs/shell-to-flang-census.md is left alone on purpose: it is a dated snapshot of 23 September with its own totals, and the row about this file belongs to that date, as the row about version-derivations-guard.sh does.
The plan asked for --na-veru, and that key costs a row in flang/scripts/run-verdict-debt.tsv, whose ceiling is full: 49 of 49, and only an overagent may raise it. run-verdict-debt-guard refused the push and named all three shortcuts. The right answer was not a row but a proof, so the imports are gone. WHAT CARRIED THE DEBT. "Lists", "String sets" and "Hash map" pulled 51 obligations into the closure that the kernel does not prove -- it counts them on the author's grid of values -- and one taken on trust. Six list helpers are written here instead (join, take, drop, dedupe and two reversals per type), and the hash dictionary is replaced by the builtin "soderzhit" over a list: the same 26 thousand lookups, but each walk happens inside the binary, not in the evaluator. NOW: utverzhdenij 16, all sixteen proved by the kernel, closure empty, and the three shortcuts run without a key. Two obligations that the kernel could not prove -- the length of a concatenation -- are not declared at all: a promise the kernel only counts on a grid is not a promise, and this file has none such. EQUIVALENCE RE-PROVED after the rewrite, on the tree of commit 5329953 with the old shell script put back beside the plan: both printed the same 17291 characters, 364 lists, 160 divergences, exit 0 for both. COST fell with the imports: 700 s instead of 1130 s, 6.6 GiB. The two ledger rows about the guard carry the new number, and so does the primechanie in the file.
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Перепись перечней, набранных руками, переехала с оболочки на план flang:
scripts/guards/hand-written-lists.sh(376 строк, два потока gawk) снят,на его месте
scripts/guards/hand-written-lists.fscript— 2755 строк, четыреплана: «Перепись» (бывший прогон без ключей), «Сверка» (
--check),«Только цели» (
--targets) и «Подлог» — проба порчи, которой у оболочкине было вовсе.
Чем доказано, что перенос верен
Оба прибора прогнаны на ОДНОМ дереве, дважды: до переписывания ввозов и после.
census)--check)Отрицательный контроль: в ведомость вписывалась строка
ПУТИ подделка/нет.md 1о месте, которого в дереве нет. Краснеют оба, обаназывают её в блоке мёртвых записей, оба печатают одно и то же число. Проба
записана ярлыком
hand-written-lists:forgeryи судит НЕ по коду (сторож красени сам по себе, долг переписи), а по имени в выводе.
Приговор доказан ядром
План идёт БЕЗ ключа
--на-веру: утверждений 16, все шестнадцать доказаныядром, замыкание пусто. Ради этого снят каждый ввоз: «Lists», «String sets» и
«Hash map» тащили в замыкание 51 обязательство на сетке значений автора и одно
на веру, а согласие на недоказанный запуск стоит строки в
flang/scripts/run-verdict-debt.tsv, где потолок полон (49 из 49). Шестьсписочных помощников написаны на месте, а хеш-словарь заменён встроенным
содержитпо списку.Что перенос изменил нарочно — четыре вещи, каждая записана примечанием
flang ioдвойка значит«кривой вызов».
перестановкой из десяти чисел и держится, пока целей ровно эти десять;
иначе берётся порядок дерева, и об этом сказано прямо.
git ls-files --othersвместо
test -eна каждое имя; «.» и «..» в имени сводятся лексически — то,что
test -eделал через ядро.6,6 ГиБ. Дороже в шестьдесят раз, и это свойство вычислителя: один зов
встроенной формы стоит около 20 мкс, а строк в обходе 775 тысяч. Сторож
по-прежнему НЕ стоит ни в CI, ни в хуке — долг, на котором он красен, записан
в
scripts/ledgers/uncalled-guards.json.Заслоны
links:checkprose-numbers-guardinventory:checkfile-extensions:checkguard-calls:checkforgery-probe:checkproved-share-tree:checklint:growthВосемь счётных примет сдвинуты прибором, а не рукой:
prose-numbers-guardназвал каждую разошедшуюся,
inventory:languagesдал новые числа.docs/shell-to-flang-census.mdне тронут нарочно: это датированный снимок23 сентября со своими итогами, и строка про этот файл относится к той дате —
так же, как строка про
version-derivations-guard.sh.