Plain-words documentation, format_time quoting fix, scratch-filesystem and release-candidate test fixes - #712

…tem and asserts what is true on each class

The row measures page-cache residency and flat Python peaks of the report/artifact path against pytest's tmp_path. On a container overlay filesystem or on tmpfs the residency property cannot hold, so the row failed there while it passes on a VM disk.

- docker/run_json_bounded_transport_test.py: a helper reads the filesystem class of the scratch directory from /proc/self/mountinfo (longest matching mount, escaped mount paths), with classes overlay, tmpfs, disk and unknown. disk and unknown keep the strict one-page bound exactly as before; overlay and tmpfs end in a pytest.skip that names the class (no tautology). Python peaks, byte and cleanup assertions stay strict on every class. The class is logged at the start and named in the failure messages. Helper rows (classification, nested mounts, missing evidence, class verdicts) are plain behaviour rows with no clock; the control rows turn red when the disk bound is loosened or when unknown is treated as non-disk.
- tests/cr150.d/core-cr150-o1w-pyjson.cmake: the registration passes -rPs so the skip reason shows in the CI log; the arguments are separated by spaces (-q -rPs -p no:cacheprovider).
- docs/ci.md: one paragraph describing the row and what it asserts on each class; adjacent prose is reflowed so the document keeps its line count (2016) that the preflight pins.

This re-authors the earlier attempt on current main and fixes its two failures (the missing space in the registration line, and the documented line count).

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

Eighteen literal cases call the real formatter with fixed timestamps, patterns
and zones and compare the returned text and its length with measured
TradingView output: the full ISO-style pattern in four zones across two
seasons, the numeric offset token alone in the same zones and seasons, a
pattern with quoted date-token letters, and a doubled apostrophe inside a
quoted literal. The test prints one PASS or FAIL line per case and returns
non-zero when any case differs.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

Takes the changes that landed on main since the branch point so the branch is main plus this change. No conflict; no hunk of the branch was altered.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…use working labels in public text

Source comments and design pages still referred to earlier work by working
labels (numbered work rounds, card-like names, phrases naming who decided,
a local-time stamp). A reader outside the project cannot follow them. Every
such mention in docs/, README.md, CHANGELOG.md, src/ and include/ now says
what was measured or decided: the instrument, timeframe, price and count are
kept, only the label is replaced by a description. Comments and prose only:
no code, string literal, identifier, macro, include or directive changes,
and no line count changes in any touched file. One content pin
(tests/pine_input_domain_gates.json) quotes a reworded comment inside its
window text and follows the new wording; its code lines are unchanged.

The sentences that carried those labels were added by earlier commits; the
doc-reverts check lists them as deletions, so they are named here:
c619f5f, e3e20cb, 0d76a09, d093560, 1e360bd, f064b1c, 665a917, 5da2668,
b97c409, cfbe677, f10f0cc, b2a578c, caa65e4, 1eef78e. Each sentence is
reworded, not dropped: the fact it states is kept without its label.

scripts/check_public_role_words.py widens the existing public-words check to
the same class (role tokens plus the working labels above). It reads every
file in scope as strict UTF-8, fails on an unreadable or undecodable file, a
failed file listing or an empty scope, and prints file and line for each hit.
scripts/test_check_public_role_words.py has one fixture per token kind, a
clean tree, an undecodable file, and a mutation row that replaces the pattern
with one that never matches and requires the failing rows to fail. The
preflight runs the test module and then the check as serial stages.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

Takes the changes that landed on main since the branch point so the branch is main plus this change. No conflict; no hunk of the branch was altered.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…h and raw directives

The new test calls the real formatter with fixed timestamps, patterns and zones
and compares the returned text and its length. It holds three kinds of rows:

- rows that characterize what the formatter returns today for literal letters,
  repeated Pine tokens, ordered mapping, raw strftime directives, unmatched
  quotes and the empty format (the rows that depend on glibc's spelling are
  compiled only where glibc is the C library);
- rows that fail today because a quoted span is expanded instead of kept
  literal, including the two measured TradingView strings, and one row whose
  output is longer than 256 bytes;
- one growth row whose output fits the old buffer but not the new initial one.

Only the test and its registration in tests/CMakeLists.txt change; no source
file is touched.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…eric offset, and size the output from the format

A pattern such as "yyyy '%Y' Z" returned the quote characters and expanded the
percent directive inside the quoted span, and a lone Z stayed the letter Z.
TradingView keeps a quoted span literal (a percent sign inside quotes is not
expanded) and writes a lone Z as the numeric offset. The formatter now

- scans quoted spans and leaves their content untouched, dropping the quotes;
  an unmatched quote keeps the previous behaviour;
- writes a lone Z as the numeric UTC offset (+hhmm) of the formatted time;
- grows its output buffer from the format itself: the previous fixed buffer was
  read past its end when the result did not fit. The growth ceiling is derived
  from the format (its length, its directive count and the widest width it
  writes), so a width a script writes is still honoured and nothing is capped by
  a fixed number; past the ceiling the formatter stops with the existing coded
  out-of-memory error instead of growing without bound.

The measured rows of tests/test_str_format_time_preservation.cpp (quoted spans,
the singleton Z, and a result longer than 256 bytes) fail on the previous
formatter by a wrong string and pass with this change; the other rows pin what
the formatter returns today. Rows that depend on glibc's spelling are compiled
only where glibc is the C library. Not covered: a libc that returns zero for an
unsupported conversion (source-inspected only), non-C locales, macOS.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…Z as the numeric offset

State, in the two coverage pages that describe pine_str_format_time, what
TradingView was measured to do and the formatter now does: a span in single
quotes is literal (a percent sign inside it is not expanded), and a single Z
outside quotes is the numeric UTC offset (+hhmm). An unmatched quote keeps the
previous behaviour; that case has not been measured.

Three lines change in two files; no line is added or removed, so line-number
anchors in the docs stay valid.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

Brings in the reviewed commit that describes past measurements in plain words and refuses working labels in public text, together with the checker that enforces it. No hunk of the merged commit was altered.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…test

Brings in the reviewed commits that make the bounded-transport row classify its scratch filesystem and assert what holds on each class. No hunk of the merged commits was altered.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…ocumentation

Brings in the reviewed commits that keep quoted spans literal and write a lone Z as the numeric offset, with the native assertions that pin the behaviour and the documentation lines that state it. No hunk of the merged commits was altered.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

Brings in the reviewed commits that let the replay golden rows accept a release-candidate version and compare timing ratios with a tolerance. No hunk of the merged commits was altered.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…s real scope

The reviewed site inventory named the growth call on the output buffer in
time_format_chunk with the wrong enclosing-scope ordinal, so the gate reported that
site as new and the declaration of the same buffer as stale. The scope is now the one
the gate derives from the file, which the preflight log of the previous commit printed.
No code and no reading change.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…ised unreviewed ceiling

The reviewed manifest now carries ten more open entries for the string sites of the
time formatter (the unreviewed ceiling in ci/cr150-source-bounds.yaml went from 1344 to
1354, the gate's designed mechanism for new open entries). The real-tree row of the gate
test pins the exact open count, so its expectation moves from 1344 + 184 to 1354 + 184.
The gate itself reports no finding and no change in its checks.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…istrations moved tests/CMakeLists.txt by

The evidence file pins the cited line of the pooc_close_fill_view_tapes guard in tests/CMakeLists.txt, and the sweep fails when the token no longer sits on that line. The two native test registrations added near the top of the same file moved the guard from line 3422 to 3424, so the cited line, and the line numbers named in the row's reason text, move by two. No other row cites that file; no code and no checker change.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>