Proc::ShellSubset::Large: the large parser, its specification and its measurements

Written with Claude by the parser session (file prefix large-subset-), 2026-11-01 and 2026-11-02. The module was called ShellSubsetLarge until the distribution moved every module under Proc:: (2026-11-02). The current documents below use the new names; the historical ones keep the names of their day.

| File | What | |---|---| | large-subset-spec.md | The specification (current): what the large grammar accepts beyond Proc::ShellSubset, the rules where dash and bash differ (A3), the tree additions, the rules for the executor at run time (A4: status 248, A5: as dash), tests and measurements. | | large-subset-state.md | The state of the union (current): the rules and which beats which, the yardsticks decided by the shells alone, the parse results, the work list, subtle and open verdicts, blatant differences; section 7: the behaviour, every accepted sample run in dash, bash and ours (the run of post-0.12, 2026-11-17; ours at exit, the modes of K5; the runs of post-0.09 and, with the programs inside Perl, of 0.10, IP1; W5, the event loop, checked over the library on Linux and what ran under Wine; ours through Run; the atomic calls). | | large-subset-logical.md | Rule E (current): where dash and bash differ, the plainly logical behaviour, by category, item by item; the earlier decisions checked against it. | | large-subset-environment.md | The samples and the run's environment (known issue X27, current): for every accepted corpus sample, whether the run-time comparison could come out otherwise in another environment (strict, almost strict, undecided), the codes, how to remake it; section 9: the six kinds of workaround and the runs with the environment varied (other cases of POSIXLY_CORRECT, the locale, SIGPIPE, /proc; fresh or private places; stand-ins for programs; fixed file times), footnotes to the totals of section 1 (EV1). | | large-subset-provenance.tsv | Its record: one line per accepted corpus sample, by snippet ID, with a short code; the header says how to read it. | | tools/large-subset-env-variants.tsv | Every sample whose result changed when the run's environment was varied (section 9 of large-subset-environment.md), by variant. | | tools/large-subset-match.tsv | The match across environments (section 11 of large-subset-environment.md, post-0.11): for every snippet where the shells or ours differ somewhere, whether dash and bash agree, with whom ours sides, how it is legalized (the rulings found), the class in each environment, and whether builtin-vs-coreutils holds; messages count nowhere. | | tools/large-subset-match-calls.jsonl | Its commands: every echo, printf, test, [, kill, pwd ours ran in those snippets, run alone by coreutils, dash's builtin, bash's builtin and ours (status, output, messages). | | large-subset-match-raw.md | The same match with messages compared as before post-0.11 (a message on standard output or in a file counted as a difference); not part of the analysis. | | tools/large-subset-env-setups.tsv | Step 2 (section 10 of large-subset-environment.md): the 2,902 samples it tried (undecided before step 3, every dependence named exactly: files, variables, standard input, arguments, cd, HOME), each with its class in two base runs and every example setup, its dependences and what each setup did. | | tools/large-subset-programs.txt | The program names the runs see (A23 (b), post-0.12): the runner makes a recorder for each name here, installed or not, and for no other name. | | tools/large-subset-match-known.tsv | Why ours differs from a shell where the match report cannot tell by itself: a decision or rule by name, or chance (section 11 of large-subset-environment.md; large-subset-runreport.pl -known). | | tools/large-subset-endmodes.tsv | Ours at exit (K5): the samples with a background job, each mode of PROC_SHELLSPAWN_END 5 times, side by side (large-subset-state.md, section 7). | | tools/large-subset-step3.tsv | Step 3 (section 12 of large-subset-environment.md): every sample undecided by its codes alone, with where its values from the environment land (classes pass, split, control, affects-shell, AI-…). | | tools/large-subset-env-control.tsv, tools/large-subset-env-affects-shell.tsv | The samples of step 3's classes control and affects-shell, run in example setups, every run in a root of its own; their commands run alone: tools/large-subset-env-control-calls.jsonl, tools/large-subset-env-affects-shell-calls.jsonl. | | tools/large-subset-env-races.tsv | The races, run many times in the same environment (make corpus-races, large-subset-repeat.pl): whether a difference of ours is a race or ours' own result. | | tools/large-subset-ai-classes.tsv | The undecided samples step 3's setups left out (93), classified by Claude reading them (AI- categories; large-subset-blocks.pl -ai). | | tools/large-subset-run-differences.tsv | The samples of the behaviour run where ours differs from both shells, or the shells differ and ours agrees with neither, or the messages are worded as neither's. | | large-subset-dialect-features.md | dash vs bash, feature by feature (current): every probed feature, blatant or subtle, with who decided it; the per-test record. | | tools/ | The measurement scripts (not run by make test; make corpus-help runs them all: ../test-pools.md, section 0): large-subset-corpora.pl (the sources), large-subset-ids.pl (the samples of some codes), large-subset-extract.pl (corpora to samples.jsonl), large-subset-scan.pl (both shells, ShellCheck, both parsers; --rx STRING tries the regex detector), large-subset-drops.pl (left-out samples with reasons), large-subset-remeasure.pl, large-subset-yardsticks.pl (the yardsticks, decided by the shells), large-subset-features.pl (the detector of dash/bash differences, and the per-test record), large-subset-verdicts.txt (verdicts kept by hand, also postponed free-form ones), large-subset-noise.txt (samples left out by hand as noise: placeholders, a|b|c alternatives, cut-off here-documents; post2-0.04), large-subset-reparse.pl (refresh the parser fields after a parser change), the run-time comparison (large-subset-runcompare.pl: -env varies the environment, -run goes through Run, -shells and -workers-from run ours alone after a change of ours; large-subset-runreport.pl: -b against a base run, -ours-from joins ours of such a run with an earlier run's dash and bash, -ours-vs compares ours of two runs; large-subset-threeway.pl, large-subset-builtins.pl) and large-subset-blocks.pl (the levels of large-subset-environment.md; -d: each undecided sample's exact dependences), large-subset-setups.pl (example setups from them, for large-subset-runcompare.pl -setup; section 10 there), large-subset-matchreport.pl (the match across environments; section 11 there), large-subset-valueprobe.pl (probes step 3's assumptions passthrough and splitglob with odd values), large-subset-repeat.pl (the same samples run many times: a race, or ours' own result), large-subset-bgjobs.pl (the samples with a background job, for ours at exit, K5), large-subset-calls.pl (marks the column CALLS of large-subset-provenance.tsv from the behaviour run, shell by shell; A23 (c)), large-subset-difffuzz.pl (differential fuzzing: random strings of the large subset in dash, bash and ours). How to run them: ../test-pools.md. | | dialect-cases/ | The probes: the examples of the page "Shell Dialect Choices" (in Proc-Spawner 0.03) (cases.txt, results.txt) and 102 more (large-subset-dialect-probes2.txt, -results2.txt), with their output in dash 0.5.12 and bash --posix 5.2.21 (run-cases.sh reruns them). | | (removed) | pages/, the sources of the checklist page "Shell Corpus Gaps" (in Proc-Spawner 0.04; removed after it with the page). Was: (../checklists/corpus-gaps.html): its item texts (gaps-items.txt) and lede (gaps-items.txt.lede), the builder build2.pl (with parts.css and parts.js, the Parts menu), and rebuild.pl, which puts new items into an already built page (used for its second and third rounds). The sources of the pages "Shell Large Subset" and "Shell Dialect Choices", and the older builder build.pl, are in Proc-Spawner 0.03. |

The base parser's documents (from an earlier session) are in ../parsers/: shell-subset-spec.md (the grammar of Proc::ShellSubset; Proc::ShellSubset::Small is the same grammar with fewer options), shell-subset-test-provenance.md (where the corpus tests in t/parser/shell-subset-corpus-tests.pl come from, and how to restore the corpora; the scripts are in ../parsers/tools/), shell-subset-corpus-stats.md (the first corpus measurements, 2026-10-29/30, pre-0.01; in Proc-Spawner 0.03). They use the old module names.