Avoid duplicate template memory and record native startup counter scopes

[seathatflowsinourveins]

Scope

Exclude automatic loading of three example/scaffold instruction templates through documented project claudeMdExcludes, and record native startup-counter scopes. The anti-pattern rule uses the client's counter rather than characters divided by four. Historical owner-reported Claude/Codex usage remains separate from the retained native Codex input-structure snapshot. Native memory-loading acceptance is pending the command center's decision.

• Base commit: d9eb68750311d5f888adec140723da174f9fe10c.
• Lane: lane:foundation.
• Owned paths: .claude/settings.json, docs/harness-defaults.md, the new template-memory decision and receipt, startup snapshot/fold receipt, and evidence registry. The three Propose excluding Claude instruction templates from nested memory #822 source files were copied byte-for-byte at6f354e93; every prior project setting is preserved. This is A14's content fold of Propose excluding Claude instruction templates from nested memory #822 into Avoid duplicate template memory and record native startup counter scopes #823, without rebase, merge or original PR closure.
• North-star action: budget actual client context while building the foundation for US-equities R&D.

SOTA sources

• Installed codex-cli0.160.1 debug prompt-input --help and its native JSON read-back: the supported renderer emits model-visible messages, with no token-usage field. The vendor's context documentation, read 2026-10-07, distinguishes startup content and later context additions.
• Owner's dated IMPROVEMENT-MANIFEST.md, C8:152-171 (2026-10-06): historical native /context display and Codex usage ranges. The source lacks session IDs and exact capture times, so these remain explicitly transcribed owner evidence.
• native-agent-stack@d9eb68750311d5f888adec140723da174f9fe10c:tests/test_adoption_docs_consistency.py:789-835: existing anti-pattern table contract. The three-file byte gate is a limited structural check, not whole-startup token measurement.
• Claude memory exclusion documentation, read2026-10-06: documented absolute-path globs, leading **/, project settings and merged arrays. Its AGENTS rules cover the scaffold import/fallback target. anthropics/claude-code@fbe20e00e2851fc01506f54f98a8f0b875af3847 (v2.1.292) CHANGELOG.md:2784 is the retained symlink-exclusion source; installed native help/version observations remain the original donor's evidence.
• native-agent-stack@d9eb6875:adoption/new-wsl/client-config-map.json:3 and tools/adoption/new_wsl_client_config.py:130-147: the map covers enumerated adoption-template inputs, not this project setting. Full owner acceptance/limits remain in the copied decision.

Evidence-class table

Claim	Evidence class	Command / receipt
Fresh Codex message/part byte and character scopes	native_proven	Native debug prompt-input, exit 0; sanitized receipt with private JSON SHA256
Claude displayed memory total 11.5k and reported Codex input ranges	historical owner-reported native evidence	Dated source transcribed; no new inference or independent usage validation
Anti-pattern table contract and registry integrity	local_integration	39 documentation tests, one skip; validate.py passes
Fresh main-checkout provider input-token total	unknown	Native renderer has no usage field; no estimate supplied
Source-preserving #822 fold and settings contracts	local_integration	Exact donor-byte comparison;64tests/one skip and structural validation
Installed Claude nested-memory exclusions	pending	CC decision and fresh native /context read-back

Local commands run

Content-fold qualification:

nice -n19 ionice -c3 python3 -B -m unittest tests.test_secret_path_guard.SecretPathGuardTests.test_project_settings_register_hook_and_deny_rules tests.test_secret_path_guard.SecretPathGuardTests.test_every_anchored_read_deny_has_a_context_mode_twin tests.test_install_claude_profile.SecretGuardProfileTests tests.test_install_claude_profile.CommittedSettingsFallbackGuardTests tests.test_install_claude_profile.PortableTopRuleTests tests.test_adoption_docs_consistency
exit0;64tests,one skip,15.371s;executed viaRTKproxy
python3 scripts/component_matrix.py --write
python3 scripts/new_host_grand_list.py --write
exit0/0
python3 scripts/validate.py
exit0;native integrity/scope only
git diff --check
exit0

Native byte comparison proves the project settings, proposed decision and receipt equal #822's source6f354e93. The original C8 receipt and all donor observed results remain unchanged; NEW fold822-20261007.json links the source and qualifies this content fold. Registry is last. Prior qualifications below remain labelled retained; no fresh model or Claude context run:

rtk proxy nice -n 19 ionice -c3 codex debug prompt-input
# exit 0; read-only main-checkout native renderer; no inference
rtk proxy nice -n 19 ionice -c 3 python3 -m unittest tests.test_adoption_docs_consistency
# exit 0; 39 tests, one skip
rtk proxy nice -n 19 ionice -c 3 python3 scripts/component_matrix.py --write
rtk proxy nice -n 19 ionice -c 3 python3 scripts/new_host_grand_list.py --write
rtk proxy nice -n 19 ionice -c 3 python3 scripts/validate.py
# exits 0; 69 components, 10,382 hashes, 4 profiles, 223 receipts
rtk git diff --check
# exit 0

The receipt preserves rounded displays and ranges, keeps deferred schemas outside usage, and never sums overlapping counters. The whole-message character labels are corrected to distinguish their isolated Skills/AGENTS text parts; invented joining separators are excluded. Raw prompt/configuration/conversation content remains private. Registry commit is last.

Decision record

docs/decisions/2026-10-06-claude-template-memory-exclusions.md is the copied proposal for the command center, including alternatives, external-path/merged-array limits, rollback and native acceptance. Other hosts remain unmeasured until their own native snapshots exist.

Host evidence

No evidence/hosts/ changes, host apply, credential-file access, model request or main-checkout edit. The command center applies any accepted project-memory policy, then a fresh native session reads an ordinary example/scaffold source and retains /context showing the three templates absent while root/blueprint instructions still load. Rollback removes only those reviewed project globs and uses a fresh session. No runtime success or token saving is inferred from repository checks.

Landing path: the co-op verifies A14's #822→#823 content fold and closes #822 with a record. This consolidated draft needs the command center's instruction/memory-scope read and ACK, then a named5f landing cue in the remaining queue. Its single landing rebase uses the hot-file protocol; there is no automatic DIRTY rebase. Host acceptance remains the configuration owner's subsequent native read-back. No close/delete/rewrite is done by this lane.

Checklist

• No GitHub Actions changes.
• No secrets printed or committed; no required secret added.
• No paid hosting or billing surface introduced.
• Peer-owned paths and worktrees preserved.