Index-captured CI failure logs and tighten triage cache integration

NOTICE

@@ -20,6 +20,12 @@ Copyright (c) 2026 EFrMG; MIT License. The permission notice is retained in
 licenses/triage-o-mator-MIT.txt. The adapter removes install-specific imports;
 it does not acquire evidence or import triage decisions.
 
+The native CI log adapter in src/reposition/ci_logs.py also adapts offline
+validation from triage-o-mator bin/_ci_logs.py, bin/_checks.py and
+bin/_acquire.py, Copyright (c) 2026 EFrMG, under the same retained MIT license.
+It removes acquisition and install imports and adds frozen bounded associations
+and raw UTF-8 segment projections.
+
 Supporting text in the README header and social preview graphics uses DejaVu
 Sans glyph outlines. Copyright (c) 2003 Bitstream, Inc. All Rights Reserved.
 Bitstream Vera is a trademark of Bitstream, Inc. DejaVu changes are in the

README.md

@@ -65,12 +65,14 @@ evidence, not automated duplicate decisions or calibrated confidence scores.
 The `cache-index`, `cache-query` and `cache-retrieve` commands work directly with
 triage-o-mator's immutable evidence cache. They provide component-aware ranked
 fragments, exact snapshot/corpus selection, source verification and progressive
-reads across all nine components. The complete JSON response has an enforceable
+reads across all nine snapshot components and separately saved CI failure-log
+segments (`ci_logs`). CI captures keep their raw segment ranges and exact check,
+run and attempt bindings; they do not establish complete job-log coverage. The complete JSON response has an enforceable
 UTF-8 byte budget. Derived indexes live beside the acquisition cache and have
 their own storage ceiling.
 
 See the [workflow and contracts](docs/cache-integration.md), including resource
-limits and an optional `bin/cache` integration patch. This API
+limits and the maintained triage-o-mator `bin/cache` bridge. This API
 keeps acquisition and approvals with the caller; real-cache and independent
 review evaluation remain the next qualification steps.
 

docs/cache-integration.md

@@ -50,7 +50,7 @@ comment, field or diff region. Its default starts at the matched unit; an explic
 `--byte-offset 0` starts at that source boundary. Follow a fragment's returned
 `continuation` parameters to read further without changing source selection.
 
-An optional [triage-o-mator bridge](../integrations/triage-o-mator/README.md)
+The optional [triage-o-mator maintained bridge](../integrations/triage-o-mator/README.md)
 provides `bin/cache search-index`, `query`, `retrieve` and `search-info`, while
 preserving existing literal search and readers when Reposition is absent.
 
@@ -61,6 +61,16 @@ match the indexed selection, even with an explicit `--db`. Without a scope,
 metadata only: it does not verify source checkpoints or payloads, and remains
 usable for inspecting a stale index before rebuilding. Its JSON explicitly
 marks both source-verification flags false; it is not a freshness claim.
+The `reposition.cache-info.v2` response gives `indexed_snapshot_count` and
+omits `indexed_snapshots` by default. Consumers of the v1 full list must opt
+into pagination. Use `--snapshot-ids-limit N` (maximum 100) to request IDs in
+sorted order, and pass the returned
+`indexed_snapshots_next_offset` as `--snapshot-ids-offset` for the next page,
+with the first page's manifest `checksum` as `--checkpoint`. Nonzero offsets
+require this checkpoint; any supplied checkpoint must match the current manifest
+checksum. Replacement therefore rejects continuation from an older index.
+A null next offset means the list is complete. The manifest still retains the
+complete checksummed set of IDs.
 
 ## Scope and authority
 
@@ -102,6 +112,7 @@ Projection version 1 retains original string slices:
 | `checks` | Check run/suite/status names, output, status, conclusion and kind |
 | `closing_issues` | One explicit relationship URL |
 | `timeline` | One event, body or available linked-issue field |
+| `ci_logs` | Separately captured native CI failure-log segments and retained check/workflow/job names |
 
 Long fields/regions split at deterministic UTF-8 byte and line ceilings, with
 small overlap. Defaults: 4,096 bytes, 200 lines and 128 bytes overlap. Chunk policy
@@ -115,6 +126,62 @@ diff offsets count UTF-8 bytes in the object itself. All ranges are half-open.
 Locators retain source record IDs where available, filename, line range, original
 field size and comment/hunk boundary. Hashes cover exact original UTF-8 slices.
 
+## Separately captured CI failure logs
+
+Default builds also discover currently saved native `ci-failure-log` captures for
+verified failed checks of the exact selected PR snapshots. Use `--component
+ci_logs` to index only these captures. An explicit ordinary component selection,
+such as `--component summary`, skips CI association discovery. No logs are fetched,
+no install code is imported, and no source cache files are changed.
+
+This is a separate searchable component, outside snapshot component completeness.
+The adapter verifies the selected summary and typed checks, derives each exact
+native `ci-logs` pointer path, and freezes its bytes and immutable packet reference
+before indexing. There is no directory scan or latest/historical fallback. The
+build rechecks the selected pointer bytes before publication; drift refuses the
+replacement and preserves the old index. Later pointer replacement or deletion
+does not change an existing index's frozen selection. Rebuild explicitly to adopt
+new captures. Packet references live in SQLite source bindings, not a potentially
+large metadata response. Metadata reports the bounded selection digest and counts.
+
+`--max-ci-logs` defaults to 1,000 (maximum 10,000). It bounds inspected PRs,
+eligible check/pointer lookups and captured packets. Capped members/checks, missing
+associations, unsupported providers, unavailable or partial source metadata,
+unavailable/partial captures, packet omissions and non-UTF-8 segments remain visible in aggregate
+coverage. Unverifiable legacy checks are unavailable for CI discovery; they remain
+eligible for ordinary snapshot-component search. The discovery pass has a separate
+128 MiB unique source-object read ceiling and respects per-object limits.
+
+Query and retrieve use only frozen immutable packet references. Each returned
+capture revalidates its packet hash and size, capture and selected snapshots,
+summary identities, raw typed checks and suites, head, check source object/index,
+check name, workflow run/attempt, job, attempt membership and segment bounds.
+A capture from another metadata snapshot is accepted only with revalidated binding
+and exact equality of PR identity/head and complete raw check and suite data;
+both snapshot IDs remain in fragment metadata. Checksums verify retained content,
+not the authenticity of external provenance or current GitHub state.
+
+Raw base64 segment bytes must decode as strict UTF-8 to be searchable. ANSI
+escapes are masked with spaces for FTS matching and excerpt placement, so
+colored words remain discoverable within short excerpts. Stored projections and cited
+excerpts preserve original ANSI/control characters, hashes and offsets; no
+normalized search-text offsets are presented as source coordinates. Unsupported non-UTF-8 segments are counted and omitted. Segment
+locators use `ci-segment-utf8`, with offsets relative to the retained segment,
+original job-log byte/line ranges and segment index. Byte ranges are half-open;
+`original_start_line`/`original_end_line` preserve the native inclusive line range.
+Excerpts also expose original job-log byte offsets. Hashes cover the exact decoded UTF-8 source slice. Name
+fields use the existing decoded JSON string coordinate space. The immutable
+packet SHA-256 URI and structural locator identify the retained raw source.
+Progressive reads stay inside one retained segment and never bridge omitted lines.
+CI fragment metadata retains provider, source repository, status, head,
+check/workflow/job names and IDs, run attempt, partial/revalidation flags and
+omissions. Partial capture, unvalidated binding and omitted regions remain separate
+from snapshot coverage and action approval.
+
+The request verification ceiling includes unique packet, checks and summary
+object bytes, including both snapshots' source dependencies when reused. Whole
+response, projection, unit and index-storage ceilings apply to these fragments too.
+
 ## Query and output policy
 
 FTS5 `unicode61` provides case folding and its default diacritic normalization,

integrations/triage-o-mator/README.md

@@ -1,93 +1,38 @@
-# Optional triage-o-mator integration
+# triage-o-mator integration
 
-This proposed patch adds four offline commands to triage-o-mator's `bin/cache`.
-The first draft PR targets current master at
-[`80cdcdd77d875c4fdcd96f3367f2f8a7d253e63a`](https://github.com/EFrMG/triage-o-mator/tree/80cdcdd77d875c4fdcd96f3367f2f8a7d253e63a).
-Earlier compatibility checks cover the `write-operations` commit
-[`f956c1299ab7f28107aa4db7943d97ce0baa4f2a`](https://github.com/EFrMG/triage-o-mator/tree/f956c1299ab7f28107aa4db7943d97ce0baa4f2a).
-The patch also supports its original base,
-[`dddc487660ffda17ba7e3626c9d9be72e29a6dce`](https://github.com/EFrMG/triage-o-mator/tree/dddc487660ffda17ba7e3626c9d9be72e29a6dce).
-The historical receipts cover the earlier patch; the refreshed patch and
-[current-master receipt](../../validation/triage-master.json) name the current
-integration base. Recheck later upstream changes before adoption.
-Existing commands work without Reposition. The patch does not alter acquisition,
-TUI notifications, ledger decisions or GitHub actions.
-
-To try the complete synthetic retrieval workflow before modifying triage-o-mator,
-run `./reposition demo --cache` from a Reposition checkout. It needs no installation
-and handles the fixture, IDs and source-verification steps automatically.
-
-This bridge supports Reposition `0.2.0.dev1` and cache format version 1. Install
-the tested engine into the Python environment used by `bin/cache`:
-
-```sh
-python -m pip install --no-deps \
-  'git+https://github.com/Univeracity/reposition.git@10bd0641b50d514984dad6dee480f140ab86ee44'
-```
-
-The bridge rejects other engine versions until compatibility is checked. Use
-this source or the matching supplied wheel: the unrelated package with the same
-PyPI name is not this project. Installation may need network access; subsequent
-indexing, queries, retrieval and metadata inspection remain offline. Reposition
-has no third-party runtime dependencies.
-
-From a clean upstream checkout:
-
-```sh
-git apply --check /path/to/reposition/integrations/triage-o-mator/cache-retrieval.patch
-git apply /path/to/reposition/integrations/triage-o-mator/cache-retrieval.patch
-```
-
-From the selected triage install:
+triage-o-mator maintains its optional Reposition bridge in
+[`bin/_reposition.py`](https://github.com/EFrMG/triage-o-mator/blob/master/bin/_reposition.py).
+Its [installation and supported-build guide](https://github.com/EFrMG/triage-o-mator/blob/master/docs/reposition.md)
+owns the pinned engine version, bundle installation and bridge configuration.
+Use that guide for the build supported by your triage-o-mator checkout.
+Reposition does not carry a second bridge implementation or installation patch.
 
-```sh
-bin/cache --expected-repo OWNER/REPO search-index --snapshot INVENTORY_SNAPSHOT --component summary
-bin/cache --expected-repo OWNER/REPO search-index --corpus CORPUS_ID
-bin/cache --expected-repo OWNER/REPO query --corpus CORPUS_ID \
-  --query 'terminal fails after suspend'
-bin/cache --expected-repo OWNER/REPO retrieve --corpus CORPUS_ID \
-  --unit UNIT_ID --checkpoint INDEX_CHECKPOINT
-bin/cache search-info --corpus CORPUS_ID
-```
+From a selected triage install, the bridge provides `bin/cache search-index`,
+`query`, `retrieve` and `search-info`. Acquisition and GitHub actions remain with
+triage-o-mator. See Reposition's [retrieval contract](../../docs/cache-integration.md)
+for source verification, selection, checkpoints and response budgets.
 
-The bridge honors install repository/host identity, requires a bound cache and
-refuses `--cache` namespace overrides. Global `--host` and `--expected-repo` go
-before the command. Other flags match Reposition's `cache-*` commands, including
-`--replace`, `--rebuild`, filters and `--cursor`. New views are stored beside the
-cache rather than consuming its acquisition budget. Installation remains optional.
-
-The patch updates `docs/evidence.md`, `docs/evidence-reference.md` and
-`prompts/prepare-analysis.md` with bounded query-then-retrieve guidance. For a TUI,
-consume the structured JSON, retain scope and checkpoint, show gaps and omissions,
-and open larger verified fragments only for selected candidates. Retrieval itself
-does not create approval or closure suggestions.
-
-Validate a trusted checkout with the patch applied:
+Validate a trusted triage-o-mator checkout with its pinned engine installed:
 
 ```sh
-python /path/to/reposition/scripts/check_triage_compat.py \
-  --checkout /path/to/triage-o-mator --output compatibility.json
+PYTHONPATH=src python3 scripts/check_triage_compat.py \
+  --checkout /path/to/triage-o-mator --pre-index-state missing-index \
+  --output /tmp/reposition-compatibility.json
 ```
 
-This uses a temporary synthetic install and upstream EvidenceCache/corpus
-publication. It checks legacy search without Reposition, missing-package behavior,
-identity guards, verified query/retrieve and unchanged authoritative cache bytes.
-It also checks `search-info` against foreign host/name/stable IDs, explicit scope,
-and literal search before/after bridge use. Metadata inspection does not audit
-source checkpoints or payloads; its JSON reports both verification flags false.
-It makes no acquisition requests. Do not run it against untrusted executable
-checkout code. Real-cache size/latency and review decisions require separate trials.
+Run this from the matching Reposition checkout. The checker creates a disposable
+synthetic install, publishes evidence and corpora through triage-o-mator's owning
+code, and checks verified retrieval, identity guards, metadata inspection,
+`pr-checks` compatibility and unchanged authoritative cache bytes without GitHub
+requests. Its default `--pre-index-state missing-package` supports the earlier
+standalone bridge trial; the maintained pinned integration expects `missing-index`.
 
-The [validation receipt](../../validation/triage-master.json) records
-the checked upstream commit, patch SHA-256 and successful compatibility results.
-It covers the pinned commit; later branch changes require another check.
+## Historical trials
 
-See the [complete retrieval contract](../../docs/cache-integration.md).
-
-The patch includes two workflow tests and a separate CI job with the pinned
-optional engine; the ordinary CI job retains the no-package path. Run checks on
-both sides whenever cache/adapter behavior changes. The
-[review and cost trial](../../docs/triage-review-trial.md) contains reproducible
-same-symptom/different-cause examples, query/retrieval costs, memory and storage
-measurements, and the boundaries of synthetic validation. Real-backlog trials
-remain the next qualification step before TUI or suggestion workflows.
+The [original integration receipt](../../validation/triage-master.json) records a
+proposal tested against triage-o-mator commit
+[`80cdcdd77d875c4fdcd96f3367f2f8a7d253e63a`](https://github.com/EFrMG/triage-o-mator/tree/80cdcdd77d875c4fdcd96f3367f2f8a7d253e63a).
+The [review and cost trial](../../docs/triage-review-trial.md) records that earlier
+engine's synthetic retrieval measurements and limitations. These are historical
+records and do not validate the current build; rerun compatibility checks when
+the engine or owning cache contracts change.