docs(changelog): record the measured fixes under Unreleased

[fajarhide]

Changelog entry for the work merged in #103. CHANGELOG.md only — no code.

Contents

Four Fixed entries (YAML block-scalar detection, find deleting its payload, the TOML filter shadowing the test distiller, cargo test recounting collapsed lines), one Removed (est_cost_usd / ModelPricing), and two Changed (the measured numbers; the determinism and latency costs, now stated rather than hidden).

One thing worth a reviewer's eye

The [0.6.1] entry currently credits the find distiller with:

strips out raw file path samples entirely and emits a highly compressed, AI-friendly key-value summary … maximizing token reduction (up to 99.8% savings)

That behaviour was the bug. It reported ~99% by discarding the file paths that were the answer to the question, and it is where the retired 97.3% all-time headline came from.

I did not rewrite the 0.6.1 entry — published history should stay as published — so the new Unreleased entry names it explicitly instead, and the old claim no longer stands as the only account of that change. The replacement is lossless prefix factoring: ~72% on the same output, every path round-tripping byte-for-byte.

Note on the numbers quoted here

They are the measured ones from #103: 58.9% net across 1,810 replayed traces, 63.6% of calls saving nothing and adding zero bytes, and a 65.3% → 58.9% drop caused by fixing the YAML sniffer — six of those points were destroyed manifests, not removed noise. Nothing in this entry is projected.

PR Auto Describe

Summary

Critical data integrity and accuracy fixes: format-safe compression now passes JSON/YAML/CSV through byte-for-byte; the YAML sniffer no longer misclassifies block scalars (fixes manifests seen as docker logs); the find distiller replaced payload deletion with lossless prefix factoring (~72% savings, nothing lost). Removed the ModelPricing dead code that couldn't know cache vs. input costs. Replaced all published benchmarks with measurements from 1,810 real traces (net 58.9% reduction, not the claimed 97.3%).

---

Key Changes

• Format gate: JSON, NDJSON, YAML, TSV, CSV bypass collapse markers entirely
• YAML sniffer: block scalar bodies (|, >, etc.) now skipped, not judged — prevents manifests misidentified as docker logs
• find/grep distillers: lossless prefix factoring replaces lossy deletion; same ~72% savings, fully reversible
• TOML filter guardrail: filter only wins if it beats MIN_REDUCTION_PCT; TestDistiller shadowing fixed
• cargo test totals: quotes runner's own summary line (cargo, pytest, jest) instead of recounting collapsed lines
• Removed: est_cost_usd, ModelPricing, _cached/_cache_creation fields (dead code, wrong numbers)
• Benchmarks: replaced all published claims with 1,810 real replay traces; actual savings: 58.9% net (was 97.3%), git 91.3%, cargo 96.8%

---

Detailed Breakdown

Data Integrity Fixes

Issue	Symptom	Fix
Format-agnostic collapse	JSON dashboard output had 14 collapse markers; jq failed	Format sniffer gates both hook points; structured formats pass through
YAML block scalar misdetection	608-line ConfigMap with HCL block scalar → 13,463 bytes as "docker logs"	Sniffer skips block scalar bodies; `
find deleting its answer	~99% "savings" = discarded file paths (the actual answer)	Lossless prefix factoring: hoist shared prefix to header, paths relative; round-trips byte-for-byte
grep per-match path repetition	Path printed per match (wasteful)	Path hoisted once per file, offsets for matches

Distiller Selection Fixes

• TOML filter shadowing: sys_build_domain was winning race but only cutting 1% (below MIN_REDUCTION_PCT); its phantom cut still discarded TestDistiller's 94% reduction. Filter now only short-circuits if it actually beat the guardrail.
• cargo test recounting: CollapseMode::Test folds ... ok lines before distiller sees them → reported 1 passed for 490 passed. Now quotes runner's summary; falls back to counting only when no summary printed.

Removed Dead Code

• ModelPricing used only .input; _cached/_cache_creation were always zero; every dollar figure was a fresh-input guess presented as session cost. Removed from stats, diff, guard::config. Unknown keys in existing config.toml still parse.

Benchmark Corrections

Claim	Actual
97.3% all-time	58.9% net (15.0 MB → 6.2 MB) — previous figure was manufactured by find deleting its payload
docker build 99.5%	37.2% (5,783 bytes from 9.2 KB fixture)
git diff 50.0%	44.6%
~40% faster TTFT	Never measured
$35+ USD/month	Used removed estimator

Real benchmarks from 1,810 execution traces on release binary with fresh HOME per invocation: git 91.3%, cargo 96.8%, cat 9.1%.

New Documentation

• Non-determinism: warm ~/.omni/omni.db causes scorer variance; same binary/input gave 1,835 bytes then 433 on identical runs (21/30 differed). A/B tests must isolate state per invocation.
• Latency cost: ~82 ms for 496-byte git status against fresh DB → ~308 ms against 97 MB DB, per hooked command.

---

Notes

• YAML sniffer fix reduced reported savings (65.3% → 58.9%) because six previously "saved" data points were actually destroyed manifests being misread as noise.
• All i18n READMEs carry the same corrected numbers.

---

Breaking Changes

• ModelPricing removed: unknown config keys still parse (backward compat), but dollar figures no longer printed in stats/diff output. Users relying on est_cost_usd output will see nothing.
• Benchmark numbers changed significantly: any docs/landing pages hardcoding old figures are now inaccurate.

Last updated: 2026-07-17 04:44:19