Bless the ledgers #10855 left red: t27b row for hslm_bwd_ptx, dupe baseline, orphan ceiling, types Range #10930

Follow-up of #10855 (merged) and #10895 / PR #10914 (split-K kernel 3). Owner's order: bless the ledgers that #10855 left red on master. The owner delegated it to the Queen to save the interactive session's tokens.

Background for the bee. #10855 added hslm_bwd_ptx.t27 and merged with several non-required ratchets red. Since then #10889 and #10913 added the rows for sar_adc, reap, ngram_hslm and others, so the only row still missing for this spec family is hslm_bwd_ptx.t27. Work from current master. Run every t27c and tri command on the Railway lab, never on a Mac. No new foreign-language code: ledgers are tool-written data, and the bless is done with the existing tri commands and tools/dupe_scan.py --bless (run, not edited). Do not touch kernel 3 or the spec itself; PR #10914 owns that. Open the PR against master as a draft with the no-auto-merge label.

Boundary

docs/reports/t27b_expectations.json
docs/reports/orphan_modules.json
docs/reports/type_conflicts.json
tools/duplicate_bodies_baseline.txt

User Scenarios

1. A reviewer opens a PR and sees the t27b-native-ratchet, Corpus ratchet, duplicate-bodies, orphan ceiling and types Range checks. After this issue they are green on master for reasons unrelated to the PR.
2. A maintainer reads the ledger and finds exactly one new row for specs/igla/coder/hslm_bwd_ptx.t27, with the status the reference really gives it (pass).
3. The orphan ceiling equals what master really has, checked first: the orphans (_enrich/mod.rs, memory/*) are master's own, not from spec(igla): hslm_bwd_ptx -- HSLM backward dX and dW as int8 tensor-core PTX written by t27 (Closes #10847) #10855.

Requirements

1. In docs/reports/t27b_expectations.json add the row for specs/igla/coder/hslm_bwd_ptx.t27 with the status the lab's tri t27b ratchet --bless records (reference passes it). No other row changes, and the cap stays at 129 unless the new row is not-pass.
2. Run python3 tools/dupe_scan.py --bless on the lab and commit only the ledger diff it writes, if there is one.
3. Check with tri mods orphan --gate on current master. If the ceiling in docs/reports/orphan_modules.json differs from what master really has, set it to the measured number; say in the PR body which files are the orphans and that they are not from spec(igla): hslm_bwd_ptx -- HSLM backward dX and dW as int8 tensor-core PTX written by t27 (Closes #10847) #10855.
4. Run tri types ratchet on master. If a Range conflict is reported, find which spec introduces it (grep -n "Range" specs/igla/coder/hslm_bwd_ptx.t27). If it comes from that spec, bless it with tri types ratchet --bless; if it comes from elsewhere, do not bless it and say which spec it is.
5. Read the failing scan log of the latest master run and list in the PR body which checks are still red afterwards and why.
6. PR body has a short "Ledgers" section: each bless and why. English only, ASCII.

Success Criteria

• grep -c "specs/igla/coder/hslm_bwd_ptx.t27" docs/reports/t27b_expectations.json prints 1
• git diff --stat origin/master -- docs/reports | tail -1 prints at least 1 file changed
• git diff --name-only origin/master | grep -v -e "^docs/reports/" -e "^tools/duplicate_bodies_baseline.txt" | wc -l prints 0
• ./scripts/tri t27b ratchet --run run.json does not print UNLISTED for hslm_bwd_ptx.t27

Requirements

• FR-001: The ledger at docs/reports/t27b_expectations.json MUST contain exactly one new row for specs/igla/coder/hslm_bwd_ptx.t27 with the status recorded by tri t27b ratchet --bless on the Railway lab; all other rows and the cap of 129 MUST remain unchanged unless the new row is not-pass.
• FR-002: The command python3 tools/dupe_scan.py --bless MUST be executed on the Railway lab and only the ledger diff it writes to tools/duplicate_bodies_baseline.txt MUST be committed, if any.
• FR-003: The orphan ceiling in docs/reports/orphan_modules.json MUST be set to the value measured by tri mods orphan --gate on current master; the PR body MUST list the orphan files and state they are not from spec(igla): hslm_bwd_ptx -- HSLM backward dX and dW as int8 tensor-core PTX written by t27 (Closes #10847) #10855.
• FR-004: If tri types ratchet reports a Range conflict, the spec introducing it MUST be identified; if it originates from specs/igla/coder/hslm_bwd_ptx.t27, it MUST be blessed with tri types ratchet --bless; otherwise it MUST NOT be blessed and the originating spec MUST be named in the PR body.
• FR-005: The failing scan log of the latest master run MUST be read and the PR body MUST list which checks remain red afterwards with reasons.
• FR-006: The PR body MUST include a short "Ledgers" section describing each bless and its reason, in English ASCII only.