Plan v2.8 batch 6: run every report family, the gates and the dry-run eraser on at least two REAL repositories that are not this one, ledger what happens, and feed the findings back as fixes. The two subjects are active local projects of the maintainer — a Python Claude Code memory plugin (70 scannable files, 1027 functions; "repo A") and a Python enforcement plugin (86 files, 1001 functions; "repo B"). Neither had ever run CodeEraser. All state created by the test (.ce/, ce-baseline.json) was removed afterwards with ce eject; both repos are identified here by shape rather than name on purpose.

Per repo, with the HEAD binaries (release build, own core): scan, dedup, dedup --check, docdup, deadcode, churn --days 14, structure, structure --split-candidates, erase (dry-run), baseline, check, trend --commits 8, doctor, eject.

Both repos measured 0/1000 on first contact. Repo A's size axis alone charged 10176‰ (one 4802-line file priced like 95 hard-line files under the quadratic extrapolation past H); repo B's charged 4325‰. Any real repository past the clamp scored identically zero — one giant file and a repo-wide disaster were indistinguishable, which is the dead-field failure this project was founded against, one representation later.

Root cause was structural, not parametric: unbounded violation mass mapped linearly onto a bounded scale and clamped. The fix replays the design from first principles — the zone curve continues C¹ LINEARLY past H (still monotone, still charging, never quadratic outside its contracted domain), and every axis charges the bounded density floor(scale·v/(v+n)) of its violation mass over its opportunity count. Full math and citations: methodology 05, wire ledger VERSIONING.md (2.17.0), design amendment size-advisory.md §A.

Scores after the fix — discrimination restored, self unchanged in ordering, CI floor re-anchored 800 → 950 at the same bite:

• Plugin-shaped repos look deader than they are. Repo A drew 43 dead-file verdicts out of 161 nodes: hook entry points are declared in JSON manifests, not imported, so the reference graph cannot see them. The verdicts are honest under the declared model (each one says "no kept in-edge and no entry flag"), and [graph] entry_globs in ce.toml is the remedy. Shipped in K step 9: when at least two files are dead and the dead share reaches half the tier — repo A's exact shape — ce deadcode now prints the entry_globs hint itself (cli/src/graph/deadcode/report.rs:78-88). A large but minority dead set stays hint-free on purpose: hinting an exemption knob at genuinely dead files would teach masking over deleting.
• erase stayed safe and honest. On both repos the dry-run plan licensed at most one row and demoted everything else to advisory with named reasons. language_unresolved on a plain .py file reads like a detection bug but is the conservative predicate doing its job — the file's LANGUAGE had hundreds of unresolved reference sites, so liveness cannot be trusted. Shipped in plan v2.25: the advisory line now carries that language's unresolved-site count (reason_detail in cli/src/erase/render.rs) — display-only, the wire reason bits are frozen.
• Refusals refuse loudly. dedup --check without a declared budget exits 2 with a one-line reason — an unjudgeable gate never passes.
• Split advisory on giants. Under the linear arm the advisor prices real seams on the giants it used to call cohesive (repo B's largest doc: recover 28560 vs cost 6950 milli). The raw milli-penalty units read awkwardly at that magnitude — ledgered as a batch-9 polish candidate, and shipped in batch 9 P15: the advisory now leads with the ROI the core's verdict is made on and the ‰ glyph came off the absolute operands (cli/src/structure/report.rs:168-181).
• Latency held. Cold scan 0.6s / dedup 2.0s on repo A; trend --commits 8 28s cold; every command answered without a window flash or a hang.