LayerFS v0.1.5 — ordinary storage, capacity, and honest limits

LayerFS v0.1.5 — ordinary storage, capacity, and honest limits

v0.1.5 ships ordinary schema-10 Init/Commit storage as the only storage path
and removes explicit compaction. New Stores use schema 10 with 4 KiB pages: exact
content-addressed admission, compression, bounded whole-file small content below
128 KiB with bounded delta chains, large-file CDC/extents, compact scoped
namespaces and authenticated pooled metadata. Canonical identity is unchanged.

What got better, measured:

• Pending-edit capacity (#116): an equal-length overwrite of committed
  content is retained as one base root plus one bounded splice descriptor — one
  64-byte descriptor instead of three 128-byte nodes. Pending workspace capacity
  moves from 5,461 to ≈29,959 spliced files at the same budget, and the
  public default-budget route now accepts 32,000 edits (32,000 × 64 B under
  the unchanged 2 MiB budget) with independent verification of all 32,000
  changed files before and after a fresh reopen.
• Distributed SDK-edit workloads: workspace-distributed-sdk-edit-500
  4.131 s → 0.666 s (0.16×) and -100 0.45× against the earlier v0.1.5
  baseline — the same #116 mechanism.
• Packed metadata (#107): pack rows 3,457 → 1,058, pack overflow slack
  −43 %, apparent Store −1.13 %, and allocated Store −1.9 % on
  store-footprint-unique-100000 (530,358,272 → 520,142,848 B). The
  self-declared ≥1.5 % allocated figure was missed numerically (+0.02 %
  allocated) and is accepted by explicit owner decision, not re-based.
• Stable wins: payload-create-* 0.79–1.02× and dedup-workspace exact/local-100 0.62–0.66× versus published v0.1.3.

What did not get better — published, not relabeled:

• One registered Tier-1 gate fails: dedup-history-unrelated-500-mixed-v2
  measured 16.107 s against unrelated-history500 < 15 s, dispositioned by
  an explicit owner waiver. Root cause: the owner-required #116 bounded pending
  representation (+1.98 s) plus the owner-accepted #107 pack coalescing
  (+0.33 s); the breach predates this campaign, and the cell is 0.89× v0.1.3.
  Residual risk: the worst-case unrelated-history tier runs ~7 % over its
  historical gate.
• 125 WARN cells; worst ratios dedup-cdc-scattered-100 3.45×,
  dedup-cross-file-unique-100 3.35×, namespace-100 3.22×.
• 139 of 198 comparable cells are ≥15 % slower than published v0.1.3,
  median 1.34×, +33.115 s total. This is context for future target
  selection, never an acceptance gate.
• The #116/#107 mechanisms add fixed per-iteration costs of +2.49 ms/exec and
  +1.46 ms/commit (#116) and ≈+0.4 ms/phase (#107).
• store-footprint-unique-100000 construction takes 5.397 s (+36 % versus
  fsync-qualified): the #107 pack-row UPDATE cost on one giant commit.
• Ordinary full157 allocated 83,951,616 B versus a Git control live
  56,197,120 B (accepted residual).
• dedup-cdc-scattered-100 leaves an ≈452 ms unattributed gap between timer
  and command window, and the slab_send_blocked_ns mechanism remains
  unattributed.
• Endurance is not qualified: the 600 s sustained proof was
  NOT_RUN_OPTIONAL under the frozen campaign declaration.

Validation: 227 registered selections terminal — 198/198 performance (182
fresh: 56 PASS / 125 WARN / 1 FAIL; 16 reused) and 29/29 proof-only (28 PASS +
1 not run) — with 182/182 fresh independent proofs PASS, every cleanup PASS, no
S0/S1, a VERIFIED_COLD namespace-100000 member, and the native gate PASS (530
tests, 118 s) on the measured tree. The campaign is inherited from issue #120;
no benchmark was re-run for this release.

Upgrade boundary: create a new schema-10 Store. Supported
schema-6/7/8/9 Stores connect without promotion; published v0.1.3/schema-5 Stores
are rejected. LayerStackStore::upgrade_format promotes a closed schema-7/8/9
Store to schema 9 only — there is no in-place promotion to schema 10 and no
downgrade. Compacted Stores from earlier builds remain readable (read-only
compatibility path). Use matching SDK/owner/daemon builds.

This is a source-only Developer Preview, not production storage. Crash and
power-loss durability are not promised. No crates.io package, prebuilt
executable or public runtime image is part of this release.

Start with the release record,
the versioned manual,
the acceptance and waivers and
every benchmark family and case.

Existing benchmark results by family

Times below are descriptive sums of the individual case timers, not campaign
wall time and not statistical speedup estimates. Positive changes mean slower
than the published v0.1.3 checkpoint, whose cache profile is undeclared. Rows
mixing fresh and reused values are marked in the release record; this table
reports all registered performance cases in the family.

Family	Performance cases	Reused	v0.1.3 seconds	v0.1.5 seconds	Change
payload_create_read	8	1	4.179260	3.454420	-17.34%
dedup_workspace_reuse	14	1	15.875100	15.301500	-3.61%
dedup_cross_file	10	2	1.402390	3.473250	+147.67%
dedup_cdc_locality	20	1	1.484610	3.510380	+136.45%
edit_length_preserving	12	1	0.082800	0.114680	+38.50%
edit_length_changing	32	2	0.230460	0.303170	+31.55%
edit_canonical_chunk_count	12	0	0.092550	0.129570	+40.00%
init_namespace	4	0	3.053450	5.559430	+82.07%
store_footprint	6	0	8.155620	13.464210	+65.09%
tiny_file_churn	20	1	8.529740	9.714270	+13.89%
namespace_mutation	4	0	0.314360	0.452180	+43.84%
directory_construction_traversal	12	1	7.629500	8.826820	+15.69%
workspace_change_locality	16	4	10.971850	15.227050	+38.78%
dedup_branch_history	20	0	47.049480	45.614220	-3.05%
git_tool_workflow	4	1	7.498040	12.683490	+69.16%
mixed_load_bearing	4	1	8.559560	9.345510	+9.18%
registered total	198	16	125.108770	147.174150	+17.64%

The five families that are faster than v0.1.3 in aggregate are exactly the ones
the #116/#107 mechanisms and the pre-existing ordinary-path wins touch; the
dedup families' large aggregate ratios compare the v0.1.5 ordinary path
(authenticated CAS + CDC + pack assembly) against pre-authentication v0.1.3
references and are published as #112 context, never as gate results. Read the
complete per-case tables.
Existing qualified results were reused; no benchmark was run for this release.