The library offers cooperative cancellation, meaningful input/output bounds, bounded font caches and page staging, and optional coarse host memory admission. These protect specific operations and ownership; they do not bound the entire managed heap or guarantee conversion of arbitrary documents. See limits and admission and caller migration.
File conversion publishes atomically after staging. Caller-owned streams stay
open and may retain an incomplete prefix after a failure; discard that output.
Stream conversion exceeding its resident content window requires TemporaryStorage;
without a provider it fails instead of creating process-temp files. File conversion
selects explicit filesystem backing. Package, staged/output I/O and owned storage
disposal and selected DOCX/PPTX font acquisition are awaited; CPU layout and
drawing remain synchronous. On .NET 10, the BCL synchronously materializes
ZIP directory entries, so even seekable input is copied through the bounded async
intake stage before ZIP parsing. Its compressed buffer stays admitted until
extraction finishes; .NET 11 keeps direct seekable intake.
MemoryBudget admits package buffers/parts and coarse parsed content-type/XML
cache capacity, with temporary parsing/normalization headroom. One existing package
scope holds retained estimates through its final readers; failed parsing refunds
attempted growth and cache hits reuse funding. XML admission is a coarse estimate,
not an exact graph/heap measurement. The shared SVG painter separately admits
coarse source-sized DOM/path/gradient working capacity before parsing and refunds
it after painting or failure. Retained vector/PDF construction resources remain
partial; render-local scratch does not cover their later lifetime. Staging retains
coarse vector descriptor/stop/pattern capacity in its existing aggregate reservation
through emission, including when content spills. The same aggregate includes
coarse page/store records, font/image/annotation references and ordinary numbering/
collection headroom; small or spilled payloads keep their metadata funding. Cleanup
refunds after descriptors and storage are released. Each page producer also holds one coarse source-sized construction reservation
through painting, snapshots and staging awaits. It refunds on page advancement or
iterator disposal; spilled content retains its independent descriptor funding.
DOCX estimates placed text/math/tables/textboxes/stories and SVG expansion; PPTX
uses the existing rendered-node traversal, node/chart source summaries, inherited
text and SVG bytes; unused master/layout definitions do not inflate it. Prepared
balloon text grows the same page reservation. Cached chart point/category spans
also contribute construction headroom: sparse indices can require more storage
than the XML size suggests. Sequential chart frames share the largest additional
window in that page reservation. Resolved workbook formula ranges also enlarge
that grant before range records and provenance projections are constructed,
including list/array replacement overlap. Sequential frames reuse its high-water
capacity; page advancement or disposal refunds it. This covers the measured range
construction gap without a per-cell ledger. Resolved axis tick expansion enlarges
the same grant before numeric lists, captions and generated painting are built.
Source axis/plot counts and radar vertices provide coarse coexistence headroom;
repeated layout/painting shares a high-water window, including workbook-range
storage. The shared SVG painter also grows that page grant before sampled
gradient output, using the observed image expansion and page SVG source weight.
Completed opacity groups and preceding operators remain covered; repeated
pictures reuse the page's high-water capacity. The builder only borrows the
reservation. Native shading keeps its existing source-sized funding. Hidden
paint does not construct or admit unused ticks. Other workbook
and effect expansion remains unproven. These estimates cover ordinary
construction overlap; unusually expanding effects/vector inputs remain unproven.
MemoryBudget also admits
PDF content staging, private SFNT/
subset programs, retained font directories, character/advance maps and kerning-pair arrays, and decoded raster planes with PNG IDAT/scanline and baseline
JPEG component/RGB construction overlap before allocation. Image/font compression
also admits actual output capacity, growth overlap and final copying. Image payloads
remain funded through emission; font-compression storage refunds after its stream
is written. GPOS right-class decoding counts before allocating exact glyph arrays.
Font directories are pre-sized from their validated table count. Collection
discovery reuses its validated directory; extraction refunds directory scratch
before constructing the selected SFNT. Retained tags/records share the existing
private-font grant. These are coarse admission estimates, not heap ceilings.
Validated MATH assembly parts use a read-only view of the font's private SFNT.
Selected/repeated assembly arrays and placements grow an equation reservation in
the existing model scope before construction. Failed preparation refunds it;
successful retained capacity stays funded through the final layout/painting user.
One source-shaped kerning construction window covers exact coverage arrays,
class maps and retained sparse/class callbacks. Sparse callbacks retain only the
needed glyph, not an entire coverage array; empty pair sets create no callbacks.
The window reuses capacity across zero-adjustment subtables and refunds after
the batches finish. Kerning sort scratch is admitted while retained pair arrays remain
funded, then refunds after sorting; failed parsing refunds program and pair storage
together. Storage refunds after
its users finish, including failure/cancellation. Crop and generated shadow/glow
planes are admitted for their invocation; recolor reuses privately owned pixels.
Embedded-font ToUnicode streams are measured and emitted with one fixed 4 KB
formatting span, without a copied mapping list or a full text/encoded payload.
Its scopes share the caller's root; they create no private allowance. Caller-owned
buffers are not charged twice. Font leases fund acquisition and copy overlap;
explicit longer-lived cache budgets keep storage admitted through eviction or
shutdown until the final reader releases it. Font-pack metadata uses the same
explicit cache budget: coarse admission precedes parsing/indexing, construction
headroom refunds after publication, and retained capacity ends after owned indexes
are cleared and pending acquisitions finish. Returned catalog values are caller-owned.
Governed operations without a cache
budget acquire private programs. Local discovery admits actual file-buffer
capacity and bounded parsing headroom through each read/parse attempt, including
collection fallback overlap, then refunds. A measured overlapping-name source
expands into 2.56 MB of retained alias keys; the default process-shared catalog
remains outside invocation admission. The opt-in
WindowsFontResolver.CreateAsync(cacheBudget, token) instead owns coarse decoded
name/index and request-cache grants until disposal, admitting construction overlap
before decoding. Its bounded program cache uses the same host budget; outstanding
leases keep their storage funded after resolver disposal. Other significant owners
remain partial. Null budget is ungoverned. Use
representative measurements to size incidental-storage and concurrency headroom.
The existing producer now owns one model scope separate from package buffers and emitted resources. DOCX models, PPTX scenes/charts and workbook records retain coarse capacity through page production, including font/staging awaits; nested package disposal does not refund a still-used workbook model. Factory construction headroom refunds after construction. Hidden-column visibility uses one lazy, fixed-domain flag array per worksheet within this headroom, independent of range width. Source-shape estimates cover ordinary model/ placement overlap. Repeated DOCX page/static-story frames now have one additional coarse preparation grant, sized from effective section stories, the largest leading body-table header and conservative column multiplicity, then grown with observed page count before copying. Nested header content contributes to the same summary. Its high-water capacity is reused across bounded candidate passes, with headroom for overlapping layouts and continuation/mapped frames; it refunds after page production. These estimates neither count every runtime object nor prove coverage of all runtime vector/resource expansion. Remaining owners stay under review. No new quota setting or private allowance is introduced.
The historical observations below pin older source/runtime identities; they are not measurements of the current release. Calling-thread allocations are allocation volume, not peak or retained memory, and exclude background threads. Diagnostic stages overlap and must not be added. Sampled heap/process peaks may miss short spikes. Font discovery, JIT, installed fonts, GC configuration, concurrency and caller buffers materially affect the host footprint. No speedup or arbitrary- document throughput claim follows from these samples.
The existing probe is built separately against C280's exact frozen library, with no library build, Office export or qualification replay. Reuse the historical fixed-portfolio baseline under its own older runtime identity. Two additional composed sources are measured in Final and review: each input gets a fresh child, one cold conversion, one warmup and three warm samples. All four repeated/staged PDFs match accepted hashes. MB below means calling-thread allocated decimal MB.
| Input | Warm Final / review MB | Median Final / review ms | Initial fonts MB | Render Final / review MB |
|---|---|---|---|---|
| Mixed box/math report | 12.295 / 12.354 | 28.560 / 29.612 | 10.630 / 10.630 | 11.497 / 11.555 |
| Ordinary picture/box footer | 2.613 / 5.711 | 14.452 / 14.478 | 1.471 / 1.475 | 2.007 / 5.026 |
Diagnostic stages overlap and are not additive; coordinated snapshots include font preparation. The probe does not count layout invocations. Source review confirms per-conversion parsed-font reuse across bounded coordinator retries, alongside existing table/related-story measurement reuse. Font preparation remains a demonstrated cost; review painting adds material allocation outside coordinated layout. No new counter/cache/quota or speculative layout rewrite is justified by these four outcomes. Three timing samples do not establish speedup, and sampled peaks may miss brief spikes. ordinary-composition-allocation-evidence.json retains runtime/input/output/environment/stage pins and the historical baseline links.
An evidence audit on November 11 retained the aggregate summaries for this and the C281 baseline, but found seven supporting reports missing. These historical observations are not a fully reproducible 0.3.0 capacity baseline.
The existing frozen C212 library (89fe008b, root a2608a43) renders the four
fixed DOCX sources in both public presets. The probe alone is built against its
pinned library; no library rebuild or visual qualification replay.
| Source (pages) | Final warm allocation (MB) / time (ms) | AllMarkup warm allocation (MB) / time (ms) |
|---|---|---|
| Report (2) | 26.319 /104.14 | 29.634 /120.34 |
| Letter (1) | 23.603 /66.74 | 26.852 /77.05 |
| Specification (2) | 29.065 /165.07 | 32.410 /140.06 |
| Held-out memo (2) | 25.119 /101.54 | 28.302 /95.58 |
Windows/.NET10.0.13, default resolver (466 discovered faces), buffer output and the default64MiB resident window. Each input/preset uses an isolated process: one process-cold conversion followed by three warm samples; table values are medians. MB is decimal. Allocation covers the calling thread; these are neither live-heap ceilings nor throughput guarantees. OS caches are unspecified and timing is noisy, so no speedup or mode ordering is claimed.
Standalone initial font-resource preparation allocates18.171-18.750MB.
Preparation diagnostics overlap; do not sum them or infer that repeated layout
is the main cost. Independent SFNT reparses are diagnostic workloads, not proof
that production repeats those parses. All eight whole/staged output identities
remain stable and equal the qualified PDFs; fourteen retained reviewed pixel
pairs and all eight failed strict gates carry. No new Office export or raster.
Evidence: artifacts/composition-c1/compound-allocation-baseline-evidence.json.
These are workload observations, not memory ceilings. Conversion budgets protect specific expansion, retention and output boundaries. Host capacity also includes runtime/JIT, resolver ownership, concurrent conversions, input/output owners and headroom. Office and PDFium validation processes are outside library measurements.
The 2026-10-05 forced-break workload uses one-character paragraphs with forced page breaks and pinned DejaVu Sans. It tests staging/scaling, not mixed-document rendering cost. Warm file-output samples reported:
| Pages | Calling-thread allocation (MiB) | Sampled managed heap peak (MiB) | Wall time (ms) |
|---|---|---|---|
| 400 | 15.96-15.99 | 13.78-16.91 | 20.9-31.6 |
| 800 | 30.69-30.72 | 14.31-18.16 | 46.8-73.4 |
| 1600 | 60.18-60.33 | 18.97-21.91 | 95.0-165.9 |
Subsequent workstation/server GC, caller-buffer, forward-only output and concurrency experiments show that allocation volume and retained deltas can stay similar while sampled process peaks vary substantially. These remain historical workload observations, not current capacity ceilings. Full environment, source, input/output identities and tables are preserved in the original operating record. Rendering checkpoint narratives remain in checkpoint history.
Use tools/NewOperatingEnvelopeFixtures.ps1 for the forced-break fixture family;
record generated hashes rather than assuming they match older archives. Build
only the allocation probe against an already qualified library with
tools/NewFrozenAllocationRuntime.ps1; do not rebuild or requalify the library
for a measurement-only change. Run its explicit-font self-test first.
Pin source/runtime identity, input/font hashes, OS, CPU, GC, output mode, warmup and iteration count. Keep fixture generation and builds outside measured processes. Mixed-document work uses the fixed portfolio, with Final/preserved and AllMarkup/Word-compatible measured separately. Time uncontended runs before accepting an optimization. Performance optimization is deferred for the preview.
Default local-font discovery admits its transient header/directory/selected-span buffers through file reads and header parsing, including overlap with a full-read fallback. Each file refunds before the next; cache hits reuse metadata without borrowing that invocation storage. Refusal remains terminal through font recovery. The default shared catalog's metadata remains outside invocation admission; use the explicit owned factory above for governed catalog retention. Windows discovery paths may yield no fonts on other platforms; use explicit portable/math resources for supported native typography and equations.
Staging also admits coarse retained page/store records, resource/annotation references and ordinary numbering/collection headroom in its existing aggregate. Small or spilled content does not remove this metadata's lifetime; cleanup drops descriptors and awaits storage disposal before refund. No extra quota or owner.