xenia-rs/audit-runs/cache-subsystem-plan/persistent-experiment.md

Persistence experiment — XENIA_CACHE_PERSIST=1 impact on matched-prefix

Setup

• Binary: xenia-rs/target/release/xrs-c10 (active C+10 build, both engines).
• Window: 50M instructions per boot.
• Flag: XENIA_CACHE_PERSIST=1; mount falls back to $HOME/.local/share/xenia-rs/cache because XDG_DATA_HOME is empty on the host.
• Comparison baseline: C+10 default-tmpdir digest instructions=50000002 imports=40465 swaps=1 draws=0 (with stable-digest md5 b8fa0e0460359a4f660adb7605e053de); event log xenia-rs/audit-runs/phase-c10-NtQueryFullAttributesFile/canary.jsonl.
• Diff tool: xenia-rs/tools/diff-events/diff_events.py --tid-map 6=1.

Boot-by-boot digests

boot	mode	instructions	imports	swaps	draws
C+10 baseline (default tmpdir)	wipe-on-boot	50000002	40465	1	0
Persist boot 1 (cold persist)	persistent	50000003	40485	1	0
Persist boot 2 (warm persist)	persistent	50000003	40485	1	0

The persistence path is +1 instruction / +20 imports over the default-tmpdir baseline. Warm boot is identical to cold boot at the digest level (no regression observed at 50M; AUDIT-053's regression was at 500M+).

On-disk cache state per boot

After persist boot 1 (cold)

Total: 7 files / 1.4 MB.

d 4096  /access      ← BUG: should be 240 B file (canary)
d 4096  /ignore      ← BUG: not in canary's cache
d 4096  /recent      ← BUG: should be 160 B file (canary)
d 4096  /d4ea4615
d 4096  /d4ea4615/e
d 4096  /69d8e45c
d 4096  /69d8e45c/9
d 4096  /69d8e45c/e
d 4096  /aab216c3
d 4096  /aab216c3/5
d 4096  /aab216c3/a
f 2400  /d4ea4615e46ee8ca.tmp
f 5784  /69d8e45ce534ffea.tmp
f 12288 /69d8e45c9355f2f8.tmp
f 12288 /69d8e45c973a5c0a.tmp
f 12288 /aab216c35ee70e0a.tmp
f 614400  /aab216c3a2c8c185.tmp
f 685464  /69d8e45c939a9dcc.tmp

Notable:

• access, ignore, recent exist as DIRECTORIES (bug #2).
• .tmp flat journals exist (cache-build path is firing).
• Hash subdirectories exist (AUDIT-054's FILE_DIRECTORY_FILE handling works).
• Hierarchical leaf files (e.g. d4ea4615/e/46ee8ca) do not exist — the .tmp → leaf rename is being silently dropped (bug #1).

After persist boot 2 (warm)

Total: 7 files / 1.6 MB (.tmp files grew).

file	boot 1	boot 2	growth	canary's equivalent leaf size
d4ea4615e46ee8ca.tmp	2400	2800	+400	400 B (d4ea4615/e/46ee8ca)
69d8e45ce534ffea.tmp	5784	6748	+964	964 B (69d8e45c/e/534ffea)
69d8e45c9355f2f8.tmp	12288	14336	+2048	2048 B (69d8e45c/9/355f2f8)
69d8e45c973a5c0a.tmp	12288	14336	+2048	2048 B (69d8e45c/9/73a5c0a)
aab216c35ee70e0a.tmp	12288	14336	+2048	2048 B (aab216c3/5/ee70e0a)
aab216c3a2c8c185.tmp	614400	716800	+102400	102400 B (aab216c3/a/2c8c185)
69d8e45c939a9dcc.tmp	685464	799708	+114244	114244 B (69d8e45c/9/39a9dcc)

The per-boot growth of each .tmp file exactly matches the byte size of the corresponding canary hierarchical leaf. Strong indirect evidence that the .tmp contains the same data as canary's leaf but is being appended on each boot instead of being renamed-to-leaf and consumed.

This is exactly AUDIT-053's "journal-style appends per boot" pattern. AUDIT-053 predicted a runtime_error regression at warm-start because the version header would go stale; in the 50M window of this experiment, that regression has not yet manifested (AUDIT-054's report had it at 500M+).

Phase A diff vs C+10 canary baseline

Command:

python3 xenia-rs/tools/diff-events/diff_events.py \
  --canary xenia-rs/audit-runs/phase-c10-NtQueryFullAttributesFile/canary.jsonl \
  --ours   xenia-rs/audit-runs/cache-subsystem-plan/persist-warm-events.jsonl \
  --tid-map 6=1

Result:

| canary_tid | ours_tid | matched | canary_total | ours_total | first_divergence_at |
|---|---|---|---|---|---|
| 6          | 1        | 102404  | 315020       | 108471     | 102404              |

Matched-prefix unchanged from C+10 baseline (102404).

Divergence event:

canary [6][102404] kernel.return NtQueryFullAttributesFile return_value=0
ours   [1][102404] kernel.return NtQueryFullAttributesFile return_value=0xC0000034

Pre-context shows both engines reach the same path query cache:\d4ea4615\e\46ee8ca byte-for-byte, but ours returns NOT_FOUND because that leaf doesn't exist on disk (only the flat .tmp does).

Conclusion

Persistence is necessary but not sufficient. The plan's Stage 1 (rename fix) is required to convert .tmp flat journals into hierarchical leaves. Stage 2 (top-level file misclassification) is required to fix the access/ignore/ recent directory-vs-file bug. Stage 3 (flip default) follows after both bugs are addressed.

Expected matched-prefix advance after all three stages: hundreds-to-thousands of events, until a non-cache divergence appears.

Reproduction commands

cd "/home/fabi/RE - Project Sylpheed"

# Clean cache for cold start
rm -rf ~/.local/share/xenia-rs/cache

# Boot 1 (cold)
XENIA_CACHE_PERSIST=1 ./xenia-rs/target/release/xrs-c10 check \
  -n 50000000 --stable-digest --out /tmp/digest-boot1.json \
  "Project Sylpheed - Arc of Deception (USA, Europe) (En,Ja).iso"

# Inspect resulting cache layout
find ~/.local/share/xenia-rs/cache -mindepth 1 -printf '%y %s\t%p\n' | sort

# Boot 2 (warm) — same command, no rm
XENIA_CACHE_PERSIST=1 ./xenia-rs/target/release/xrs-c10 check \
  -n 50000000 --stable-digest --out /tmp/digest-boot2.json \
  "Project Sylpheed - Arc of Deception (USA, Europe) (En,Ja).iso"

# Boot 2 with Phase A event log
XENIA_CACHE_PERSIST=1 ./xenia-rs/target/release/xrs-c10 exec \
  -n 50000000 --quiet \
  --phase-a-event-log xenia-rs/audit-runs/cache-subsystem-plan/persist-warm-events.jsonl \
  "Project Sylpheed - Arc of Deception (USA, Europe) (En,Ja).iso"

# Diff
python3 xenia-rs/tools/diff-events/diff_events.py \
  --canary xenia-rs/audit-runs/phase-c10-NtQueryFullAttributesFile/canary.jsonl \
  --ours   xenia-rs/audit-runs/cache-subsystem-plan/persist-warm-events.jsonl \
  --tid-map 6=1

State at session end

~/.local/share/xenia-rs/cache/ is left in its post-boot-2 state (7 .tmp files, ~1.6 MB, 3 directories that should be files, 7 empty hash subdirs). Next session should rm -rf ~/.local/share/xenia-rs/cache before Stage 1 work to start from a clean slate.