Phase 12: cross-cutting documentation

Phase 12 of PLAN-bat-population.md. Read its phase 12 paragraph and the "Constraints that apply throughout" section, in particular the rule that every qemu-img behaviour claim is either cited to instar-testdata/qemu-sources/ or measured with the static builds, and the rule that code and user documentation do not cite this plan's phase, step or decision numbers.

This phase writes no code. It carries the documentation that belongs to no single phase, after phases 1 to 11 each updated the pages their own change touched.

Planning effort: medium. The master plan allows it, and the survey agrees: there are no design decisions here, the shape is a set of factual corrections to five named pages. The one part that is not mechanical is the pair of quirks entries, which assert what qemu-img does and so must be measured before they are written rather than reasoned about.

Review effort: medium, with one exception. A wrong claim in docs/quirks.md is expensive: it is the page the next person trusts when deciding whether a divergence is known, so a divergence described backwards there is worse than one left undocumented. The review that matters is of the two new quirks entries against their transcripts, not of the roster edits, which are checkable by grep.

What the survey found

The master plan's phase 12 paragraph was written before phases 1 to 11 executed. It names five targets; the survey found one of them needs nothing, one needs a different edit from the one described, and one of the three quirks it lists is already written. File and line references are against 41f74d40.

The per-phase constraint held, and that is the finding

The master plan requires every phase that changes user-visible behaviour to update the documentation describing it in the same pull request, with phase 12 for the cross-cutting pages only. That held. Counting mentions of the fixed subformat:

Page	Owner	subformat=fixed mentions
docs/create.md	phase 7	7
docs/convert.md	phase 8	3
docs/measure.md	phase 9	3
docs/quirks.md	phases 7-9	6
docs/format-coverage.md	phase 12	0
docs/format-internals.md	phase 12	0

docs/create.md mentions preallocation 20 times; the output side of docs/format-coverage.md mentions it zero times, its only occurrence being in the resize table. The staleness is exactly where the master plan predicted it would be, which is a real result and worth stating: phase 12 is a documentation phase with a genuine, bounded backlog rather than a sweep looking for work.

docs/format-internals.md claims dynamic-only output and a flat 32 MiB block size

Its vhdx section opens:

Microsoft VHDX Virtual Hard Disk v2 (Hyper-V). Supported:
- Dynamic VHDX: BAT-based block allocation with 32 MiB blocks (input,
  output, check)

Both halves are now false:

• Fixed output is missing. create -f vhdx -o subformat=fixed landed in phase 7 and convert -O vhdx --subformat fixed in phase 8, and measure answers for a fixed target since phase 9. The word "fixed" does not appear in the file.
• "32 MiB blocks" is right only for one size band. vhdx::default_block_size in src/crates/vhdx/src/lib.rs returns 64 MiB above 32 TiB, 32 MiB above 100 GiB, 16 MiB above 1 GiB and 8 MiB at or below 1 GiB. 32 MiB is the answer for the 100 GiB to 32 TiB band only. docs/convert.md already states the scaled rule correctly, so the two pages disagree with each other today.

The rest of the section -- the CRC-32C, metadata, overlap-rule and region-cap prose -- is current and is not to be rewritten.

docs/format-coverage.md names no VHDX subformat on the output side

Two places:

• The "Conversion Output Format Support" table's vhdx row says "Dynamic VHDX, configurable block size (1MB-256MB ...)". Its block size clause is correct, including the scaled default; its subformat claim is not.
• The "Output side" prose under the parity tables names vmdk's three subformats ("monolithicSparse, streamOptimized and monolithicFlat") but names no VHDX subformat, and says nothing about the no-op preallocation accepted on vmdk, vpc and vhdx since phase 3.

The vhdx | ✓ cell in the create/convert-output/dd-output roster table is still accurate and does not need a subformat column.

ARCHITECTURE.md's creatable roster omits monolithicFlat and fixed

Creatable target formats are raw, qcow2, vmdk (monolithicSparse and
streamOptimized), vpc (dynamic and fixed) and vhdx (dynamic).

Two omissions: vmdk gained host-only monolithicFlat in phase 3, and vhdx gained fixed in phase 7. Note that vpc already reads "(dynamic and fixed)", so the sentence's own form shows what the vhdx entry should look like.

The master plan frames this target as "ARCHITECTURE.md if the shared layout function changes the component description". That trigger does not fire: the phase 5 shared layout function did not change how the component is described. What is stale is a factual roster, which is a smaller and different edit. ARCHITECTURE.md is loaded into every session, so this phase corrects those two entries and adds nothing.

docs/quirks.md is missing two divergences and already has the third

The master plan lists three for this phase. Measured against the tree:

Divergence	State
convert's in-place fixed layout vs qemu's default	already written
-o log_size	missing
dynamic block_state_zero	missing

The first is covered: docs/convert.md states the fixed output is the layout of qemu-img create -o subformat=fixed,block_state_zero=off and points at the "VHDX partial last block" section of docs/quirks.md, which exists and discusses the block_state_zero layouts at those sizes. Phase 8 did this work. Nothing to add.

The other two are real and undocumented:

• log_size is accept-ignored. parse_create_o_options (create) and parse_o_options (measure) in src/vmm/src/main.rs both match ("vhdx", "log_size") to a bare accept-ignore arm, with the comment "log_size stays accept-ignored". instar writes qemu's default log size whatever the option says. The only occurrence of log_size in docs/quirks.md today is inside a pasted qemu-img create transcript, which is not a documented divergence. convert does not parse -o at all -- it takes --subformat -- so this divergence is create and measure only, and the entry should say so.
• block_state_zero is parsed for a dynamic image and then not read. parse_create_o_options parses it for both subformats deliberately, because qemu-img rejects a non-boolean value for both, with the comment "Only a fixed image reads it". So create -f vhdx -o subformat=dynamic,block_state_zero=off is accepted and changes nothing. parse_o_options accept-ignores it for measure entirely, which is correct and not a divergence: a fixed target is dense either way, so the measured numbers do not depend on it.

What phase 12 must measure, not assume. Both entries above describe instar. Neither says what qemu-img does, and the plan deliberately does not guess: whether qemu-img honours log_size by writing a different log region, and what it does with block_state_zero=off on a dynamic image, are the facts the entries turn on. They are measured with the static builds in step 12b before a word of either entry is written.

Corrected at source

The master plan's phase 12 paragraph and the docs/plans/index.md row are corrected in this phase's first commit, so the next reader does not re-derive the above: CHANGELOG.md is recorded as already complete, the ARCHITECTURE.md trigger is restated as a roster correction, and the convert-layout quirk is marked as written in phase 8. A later step does not need to redo this.

Decisions

1. CHANGELOG.md is out of scope. The survey found every phase from 7 to 11 wrote its own [Unreleased] entry, including fixed create, fixed convert, fixed measure, the cross-version baselines and the fuzzing. There is nothing to add, and the alternative -- rewriting correct entries into a single summary -- would lose the per-change detail a reader of the released changelog wants.

2. ARCHITECTURE.md gets a two-entry roster correction and nothing else. Not the component-description review the master plan's wording implies. The page is loaded into every session, so a documentation phase is exactly the wrong moment to let it grow.

3. The two quirks entries are measured before they are written, and each cites the qemu-img version it was measured with. This is the master plan's standing constraint, and it is the reason 12b is a separate step at a heavier model than the roster edits: the cost of a wrong entry here is paid by whoever next triages a VHDX divergence.

4. An audit step runs first and changes nothing. 12a enumerates every documentation claim about VHDX output capability, VHDX block size, and the creatable/convertible format rosters, across docs/, README.md, AGENTS.md, ARCHITECTURE.md and .claude/skills/. The edit steps then work from that list.

This is the decision most likely to be argued with, because the survey already found the five pages that are wrong and an audit looks like re-doing it. The argument for it is the one phase 11 learned three times over: the survey checked one axis -- the words "fixed" and "preallocation" -- and found the per-phase pages clean on that axis, which is not the same as clean. The "32 MiB blocks" claim is the proof. It was not found by looking for "fixed"; it was found by reading the section that the first finding pointed at, and nothing in the survey rules out a third page carrying the same number. A grep for the shape costs one cheap step and converts "the five pages I found" into "the pages that make this class of claim". If 12a's list is just the survey's five, that is a one-line result and the phase proceeds unchanged.

1. docs/format-internals.md gains the VHDX layout facts but does not duplicate docs/convert.md. The per-format page says what the fixed layout is -- every payload entry FULLY_PRESENT at its fixed offset, the data start, zero blocks left as holes -- and points at docs/convert.md and docs/create.md for the command surface. Two pages describing one option is how they come to disagree, which is the defect this phase is fixing.

2. No new tests, fixtures or code. The only executable output is 12b's measurement transcripts, which are evidence for the prose and are recorded in the execution notes rather than committed as a fixture.

Step plan

12a runs first; 12b is independent of it and may run at the same time. 12c, 12d and 12e all consume 12a's list and each run pre-commit run --all-files, so they run in sequence rather than together -- two steps running that gate in one worktree trip over each other's work.

Step	Effort	Model	Isolation	Brief for sub-agent
12a	medium	sonnet	none	Audit only, change no file. Enumerate every documentation claim in this repository about (i) which VHDX subformats instar writes, (ii) the VHDX block size or its default, and (iii) the roster of creatable / convert-output formats. Search docs/, README.md, AGENTS.md, ARCHITECTURE.md and .claude/skills/. Grep for the shape, not one word: 32 MiB/32MB/32 MB near vhdx, [Dd]ynamic VHDX, subformat, block.size, monolithic, [Cc]reatable, preallocation, and the format-roster tables. For each hit report file, the claim in one line, and whether it agrees with vhdx::default_block_size in src/crates/vhdx/src/lib.rs and with the subformats create and convert actually accept (read parse_create_o_options and the --subformat handling in src/vmm/src/main.rs). Output one markdown table. Do not fix anything, and do not report a claim as wrong without naming the code that contradicts it.
12b	high	opus	none	Measure, then write two new docs/quirks.md entries. Measure first, with the static per-version qemu-img builds in instar-testdata (see the qemu_pins helper and docs/testing.md for how the pinned builds are invoked; do not use the host qemu-img): (1) does qemu-img create -f vhdx -o log_size=<n> produce a different file from the default, and if so what differs -- compare at two sizes and at least two log_size values, including one below and one above qemu's default 1 MiB; (2) what does qemu-img create -f vhdx -o subformat=dynamic,block_state_zero=off produce, compared with block_state_zero=on and with the option absent. Then write the entries. instar's side is already established and is not to be re-derived: parse_create_o_options accept-ignores log_size and parses block_state_zero for both subformats while only a fixed image reads it; parse_o_options (measure) accept-ignores both; convert takes no -o, so it is outside both entries. Follow the house style of the existing VHDX entries in docs/quirks.md -- a heading, what each tool does, the measured transcript, and the consequence for a user -- and cite the qemu-img version each measurement used. Do not state what qemu does from reasoning; if a measurement cannot be made, say so in the entry and in your report rather than filling the gap. Record every transcript in your report.
12c	medium	sonnet	none	Rewrite the vhdx section's "Supported" list in docs/format-internals.md and add a short fixed-subformat paragraph. The list must name dynamic and fixed, and must not state a single block size: the default is size-scaled, per vhdx::default_block_size (64 MiB above 32 TiB, 32 MiB above 100 GiB, 16 MiB above 1 GiB, 8 MiB at or below), and the configurable range is 1 MiB to 256 MiB. The new paragraph says what the fixed layout is -- every payload BAT entry PAYLOAD_BLOCK_FULLY_PRESENT at its fixed offset, all-zero blocks left as holes so the file stays sparse, file length the data start plus the virtual size -- and refers to docs/create.md and docs/convert.md for the command surface rather than repeating the options. Leave the CRC-32C, metadata, overlap-rule and region-cap prose alone: it is current. Work from 12a's list: fix every docs/format-internals.md item on it, not only the ones named here. Do not cite a source line number anywhere in docs/ -- the docs-line-references pre-commit hook rejects it, in the file.rs:NNN form and in prose; name the function or constant instead. Run pre-commit run --all-files and report the hook list.
12d	medium	sonnet	none	Correct the output side of docs/format-coverage.md. The "Conversion Output Format Support" table's vhdx row must name dynamic and fixed (keep its block-size clause, which is correct); the "Output side" prose that already names vmdk's three subformats must name the VHDX subformats too, and must state that preallocation=off|falloc|full is accepted and is a no-op on vmdk, vpc and vhdx while metadata is refused -- which is qemu-img's own behaviour, so it is parity rather than a divergence. Do not add a subformat column to the create / convert-output / dd-output roster table; its vhdx | ✓ cell is accurate. Work from 12a's list and fix every docs/format-coverage.md item on it. No source line numbers in docs/ (the docs-line-references hook). Run pre-commit run --all-files.
12e	low	sonnet	none	In ARCHITECTURE.md, correct the one sentence beginning "Creatable target formats are": vmdk must name monolithicFlat alongside monolithicSparse and streamOptimized, noting it is written on the host, and vhdx must read "(dynamic and fixed)" as vpc already does. Change nothing else in the file and add no new paragraph -- it is loaded into every session and is an index, not a reference. If 12a's list names other ARCHITECTURE.md or README.md claims, report them and stop rather than fixing them: that is a scope decision for the management session. Run pre-commit run --all-files.

Risks and mitigations

• A measured claim in 12b is wrong, or is measured with the host qemu-img instead of a pinned build. This is the expensive failure: it lands a confident, wrong statement on the page people trust. Mitigation: the management session re-measures one case from each of 12b's two entries with a pinned build before committing, and rejects the step if a transcript is missing or its version is unstated. The brief names the pinned-build requirement explicitly because the host binary is the path of least resistance.

• 12b cannot measure something and fills the gap by reasoning. A plausible sentence about qemu is indistinguishable from a measured one once it is on the page. Mitigation: the brief says to report the gap instead, and the definition of done requires a version string beside each claim, which a reasoned claim cannot produce.

• The documentation phase grows into a rewrite. Five pages of factual corrections is a small change; docs/quirks.md alone is over 5,000 lines and invites tidying. Mitigation: each brief states what not to touch, 12e is told to stop and report rather than widen, and the definition of done counts changed files.

• The docs-line-references hook rejects an edit late. It covers docs/ excluding docs/plans/, and since de08bfa6 it catches prose forms -- "(line NNN)", "lines NNN-NNN", "at line NNN" -- as well as file.rs:NNN. The prose half is the trap: it is newer than most of this plan's phases, so a habit formed earlier in the plan passes review and fails the gate. Mitigation: both editing briefs name it and say to name functions and constants instead; it runs in pre-commit, so the gate catches it before review rather than in CI.

• Two of 12c, 12d, 12e run at once and collide in the gate. They touch different files but share one worktree and all run pre-commit run --all-files. Mitigation: the step plan says they run in sequence, and says why.

Definition of done

Each item is a command and its expected result, run from the phase worktree.

• grep -n '32 MiB blocks' docs/format-internals.md matches nothing.
• The vhdx section of docs/format-internals.md contains both dynamic and fixed, and grep -c 'subformat' docs/format-internals.md is greater than zero.
• No page states the VHDX default block size in a way that disagrees with vhdx::default_block_size: every documentation hit for 32 MiB, 32MB or 32 MB within a VHDX context either states the full scaled rule or names the size band it applies to. Checked against 12a's list, with each item marked fixed or deliberately left.
• The "Conversion Output Format Support" vhdx row in docs/format-coverage.md names both subformats, and the output-side prose names the VHDX subformats and the no-op preallocation behaviour.
• ARCHITECTURE.md's "Creatable target formats are" sentence names monolithicFlat and names vhdx as dynamic and fixed, and git diff --stat origin/develop -- ARCHITECTURE.md shows no more than that sentence's lines changed.
• docs/quirks.md has one entry for log_size and one for dynamic block_state_zero; each states what instar does, what qemu-img does, and the qemu-img version the latter was measured with; and each measurement's transcript appears in the execution notes.
• grep -n 'log_size' docs/quirks.md matches outside the existing pasted transcript -- that is, the word appears in prose of its own.
• No new docs/quirks.md entry claims convert accepts -o.
• pre-commit run --all-files passes, and the report names every hook that ran including docs-line-references.
• CHANGELOG.md is unchanged by this phase: git diff --stat origin/develop -- CHANGELOG.md is empty.
• No file outside docs/, ARCHITECTURE.md and docs/plans/PLAN-bat-population*.md is changed: git diff --name-only origin/develop lists nothing else, and in particular no file under src/ or tests/.
• No documentation this phase writes cites a plan phase, step or decision number, or a PLAN-*.md file.
• 12a's audit table is recorded in the execution notes, and every row is marked fixed, already-correct, or deliberately left with a reason.

Execution notes

Implemented 2026-11-11 (+1100), in one session, on branch bat-population-phase-12-docs from 41f74d40.

Back brief answers

Question	Answer
Run 12a, or go straight to the survey's five pages?	Run it as planned
Is CHANGELOG.md really finished?	Yes, leave it alone
May 12b run the pinned static qemu-img builds?	Yes
May this session push and open the pull request?	Yes

The last answer is why this phase does not stop at the commit.

Step order, and the step that was added

12a and 12b ran at the same time: 12a changes no file, so it could not collide with 12b's write to docs/quirks.md. 12c, 12d, 12e and 12f ran one after another, because each runs pre-commit run --all-files in the one shared worktree and two of those in flight trip over each other.

12f is not in the step plan above. The management session added it after 12a reported, because the audit found stale claims on two pages the survey had never opened -- docs/guest-architecture.md and docs/configuration.md. Both are cross-cutting rather than per-phase pages, so they fall inside this phase's remit, and the four corrections are factual one-liners rather than rewrites. It is a separate step so that the scope growth is visible in the history instead of arriving inside an unrelated diff.

Decision 4 was worth its step

The decision argued for an audit on the grounds that the survey had checked one axis -- the words "fixed" and "preallocation" -- and that nothing ruled out another page carrying the stale 32 MiB default. That is exactly what it found: docs/guest-architecture.md gives the VHDX writer's default block size as 32MB, and no search for "fixed" would have reached it, because that page does not use the word.

It also found a claim worth more than the rest of the phase put together: the same page said the non-qcow2 sparse formats reject non-off preallocation. They accept it. That one predates this plan, and it qualifies the phase's own headline finding -- the per-phase constraint held for the pages each phase owned, not for every page its change made wrong.

12a's audit table

Rows the audit found wrong, and what became of each:

File	Claim	Outcome
docs/format-internals.md	creatable: vhdx dynamic only	fixed, 12c
docs/format-internals.md	creatable: vmdk sparse + streamOptimized only	fixed, 12c
docs/format-internals.md	"Dynamic VHDX ... 32 MiB blocks"	fixed, 12c
ARCHITECTURE.md	creatable roster omits monolithicFlat and fixed	fixed, 12e
docs/format-coverage.md	convert-output vhdx row: dynamic only	fixed, 12d
docs/format-coverage.md	VHDX note: dynamic output, one block-size tier	fixed, 12d
docs/format-coverage.md	output-side prose names no VHDX subformat, no preallocation	fixed, 12d
docs/guest-architecture.md	VHDX writer "default 32MB", dynamic only	fixed, 12f
docs/guest-architecture.md	non-qcow2 formats "reject non-off preallocation"	fixed, 12f
docs/guest-architecture.md	"vhdx default block_size" a live writer divergence	fixed, 12f
docs/configuration.md	-O FORMAT offers "(raw, qcow2)"	fixed, 12f

Rows the audit checked and found correct, so deliberately left: the create / convert-output / dd-output roster table and the resize rows in docs/format-coverage.md; docs/dd.md's block-size tiers; AGENTS.md and README.md, which name write formats without claiming a subformat; docs/create.md, docs/convert.md, docs/measure.md, docs/resize.md, docs/info.md, docs/quirks.md and docs/testing.md; docs/commentary/reading-order.md; and docs/guest-architecture.md's resize and dd-output claims.

Two rows the audit could not resolve and the management session then settled. docs/configuration.md's -O cell was resolved by its own section heading, "Convert Options", which makes it instar's flag table rather than an integration wish-list, and then against parse_output_format, which accepts five formats and lists all five in its error message -- so it was fixed in 12f. docs/usage.md's format list is an OpenStack "must support" set rather than a roster of what instar writes, so it is left alone.

Measurements

Every qemu-img claim this phase added was measured against the pinned static builds under instar-testdata/qemu-img-binaries/x86_64/. The host qemu-img was never used. Transcripts are in this session's scratchpad under step-12b/transcripts/, covering log_size and dynamic block_state_zero at 10.2.0 and 6.0.0, their structural comparisons at 16 MiB and 4 GiB, the edge cases, and a sweep of all 80 builds (sweep-all-versions.txt, 80 versions x 9 probes).

log_size is honoured by qemu-img: the header's LogLength and the bat and metadata region offsets move with it, carrying the file length, and nothing else changes. A value that is not a whole MiB is rounded up rather than refused; log_size=0 is refused and leaves a zero-length file.

block_state_zero=off is honoured by qemu-img on a dynamic image, which this plan had not assumed either way: every payload BAT entry becomes PAYLOAD_BLOCK_NOT_PRESENT instead of PAYLOAD_BLOCK_ZERO, and nothing else changes. instar accepts the option and ignores it for a dynamic image, so this is a real divergence rather than a shared no-op.

The management session re-measured one case from each entry independently on 6.0.0 and reproduced both, and separately measured the preallocation parity that 12d and 12f assert, because the brief had asserted it without evidence. That measurement is worth recording for the next person, because byte comparison is the wrong oracle and says so loudly: every qemu-img create writes fresh random identifiers, so two creates with the same mode differ too. With identical output filenames and the per-creation fields accounted for, off, falloc and full are indistinguishable -- vmdk differs only in its CID, vpc in its unique ID and the footer copy of it, vhdx in the two header write GUIDs and one metadata GUID, and all three VHDX modes compare structurally equal through tests/helpers/vhdx_structure.py. metadata is refused on all three formats.

Definition of done, extended

The definition of done above was written against five pages, and the audit found seven. Four items were added when 12f was added, and they are marked as such in the report: a criterion written after seeing the work is weaker evidence than one written before it.

• docs/guest-architecture.md states no single VHDX default block size, and names both subformats for the VHDX writer.
• docs/guest-architecture.md does not say the non-qcow2 sparse formats refuse a non-off preallocation mode.
• docs/guest-architecture.md does not list a vhdx item among the writer divergences carried as per-case skips.
• docs/configuration.md's -O FORMAT cell names all five output formats parse_output_format accepts.

Corrections made at source

The master plan's phase 12 paragraph and the docs/plans/index.md row were corrected in this phase's planning commit, before any step ran: CHANGELOG.md recorded as already complete, the ARCHITECTURE.md trigger restated as a roster correction, and the convert-layout quirk marked as written by phase 8.

Review corrections to step output

Each editing step's output needed at least one correction, and all but one were the same shape: a sentence that would have created a fresh disagreement between two pages, which is the defect this phase exists to remove.

• 12c dropped the power-of-two constraint from the block-size range that docs/create.md states, omitted check from the fixed VHDX line where docs/check.md documents check accepting fixed images qemu-img calls truncated, and named measuring while linking only create and convert.
• 12e and 12f each left one line past the file's wrap width after an edit, rewrapped by the management session.

Two steps reported that a claim came from their brief rather than from the code -- 12c the block-size range, 12d and 12f the preallocation parity and the five output formats. Both flags were warranted, and both gaps were the brief's rather than the step's. The parity claim in particular was asserted without evidence by the management session and then had to be measured.

Back brief

Answer before any step runs:

1. Is decision 4 worth its step? 12a is a cheap audit that may return only the five pages the survey already found. Run it, or go straight to the edits on the survey's list?
2. Is decision 1 right that CHANGELOG.md needs nothing? The survey found a per-phase [Unreleased] entry for every phase from 7 to 11. A reader of the release notes gets five entries about fixed VHDX rather than one. Leave them, or add a summary entry?
3. Does 12b need a pinned-build run at all, or is there an existing recorded measurement of qemu's log_size and dynamic block_state_zero behaviour that it should cite instead? If the answer is that it must measure, confirm the session may run the static builds.
4. May this session open the pull request when the phase is done, or does it stop at the push?

Out of scope

• CHANGELOG.md (decision 1).
• Any code, test or fixture change. A documentation phase that fixes a bug it found is a phase that was not reviewed as a bug fix; anything 12a or 12b turns up that is a defect rather than a stale sentence is filed as an issue and named in the execution notes.
• The per-phase pages -- docs/create.md, docs/convert.md, docs/measure.md, docs/testing.md -- except where 12a finds one of them carrying a claim in the audited class. They were kept current by their own phases.
• The docs/quirks.md entries for divergences this plan did not create, and the existing "VHDX partial last block" section, which is phase 8's and is correct.
• Anything in the master plan's Future work list, including block_state_zero for dynamic VHDX as a feature (this phase documents that it is accepted and ignored, which is not the same as implementing it) and log_size being honoured.

📝 Report an issue with this page