epic: our own model -- IGLA CODER, HSLM in t27 and the harness, one 27-part map #10024

Goal

One map of what exists, and what does not, before the project has its own model: IGLA CODER (the code model in gHashTag/igla-coder-gpu), HSLM in t27 (the ternary language model that was trained in Zig), and the harness that has to compile, run and show them. Replaces the scattered state in #1032 (closed for lack of close criteria), #1037-#1040, #7778, igla-coder-gpu#28/#38.

The map is a t27 spec, specs/igla/own_model.t27: 27 cells through the T27F topic workflow (specs/t27f/topic_engine.t27 earned_status, checked by specs/t27f/topic_reel.t27). A cell's status is earned from measurements (source, test pass, review, live), never chosen. A cell's percent is a separate estimate with its basis named in the spec's SOURCES.

State on master, 2026-10-18

The values come from specs/igla/own_model.t27 at 82bd5f4 (#10428).

4 published, 22 verified, 1 draft, 0 gaps. Percent, equal weight per cell (a convention, not a measurement): overall 62.

The IGLA board shows the map. GET https://board-production-2dec.up.railway.app/state has a live job own_model with the message "3 published, 22 verified, 1 drafts, 1 gaps; 61%" and 27 cells. It is the event t27c gen-json specs/igla/own_model_board.t27 printed on the Railway lab, posted 2026-10-18 09:10:27 UTC.

Review: #10164, #10179, #10186, #10185 and #10270 were all merged by auto-merge before the reviewer bee approved them. #10273 took back #10270's unbacked claim. #10275 put the seven DELIVER cells to review again, labelled no-auto-merge. REVIEW_ID names review 5561941493, the bee's APPROVE of 014a34d. On the final head 24e9f71 the bee's model approved again, but the bee posted NEEDS_PERSON because the head changes two .trinity/seals/ files (t27c seal output from the lab). It merged over that hold, under the owner's standing instruction; the reason is on the PR.

Parts

Closing the epic needs cell 26 (the model computing from DDR3, reviewed).

Boundary

• In: specs/igla/own_model.t27 and the sub-issues below. Each sub-issue changes .t27 only, and updates the map's measurements and tests in the same PR.
• Out: no new hand-written non-t27 code; no model-quality claim beyond the numbers in the spec's SOURCES; nothing posted externally.
• A cell's status changes only when its measurement changes; its percent only with a named source.

Close criteria

Every cell in specs/igla/own_model.t27 is VERIFIED or PUBLISHED, measured on the lab, and the IGLA board shows the map from the spec.

Sub-issues

• the map spec itself
• Implement part 1 of 1 of the empty bodies in specs/nn/hslm.t27 (4 functions) #7778 forward pass bodies in specs/nn/hslm.t27
• make the nine specs/igla/coder/*.t27 compile
• revive the board loop: igla: repost the own_model board event from master on a Railway cron (Closes #10321) #10324 runs the Railway cron own-model-board (17 */6 * * *), which reposts gen/json/igla/own_model_board.json from master; the 2026-10-18 12:21 UTC fire returned ok and the board shows "3 published, 22 verified, 1 drafts, 1 gaps; 61%" (igla-coder-gpu#38, the wider Queen executor, stays open)
• spec(igla): own_model -- take back the REVIEWED claim #10270 merged without a review #10272 take back the unreviewed REVIEWED claim (PR spec(igla): own_model -- take back the REVIEWED claim #10270 merged without a review (Closes #10272) #10273, merged)
• spec(igla): own_model -- the seven DELIVER cells reviewed, 3 published / 22 verified #10269 the seven DELIVER drafts reviewed (7, 8, 16, 17, 18, 25, 27), PR spec(igla): own_model -- put the seven DELIVER cells to review, auto-merge off (Closes #10269) #10275, merged
• cell 9: tern-tc-500m trained, carded, published (Hugging Face playra/tern-tc-500m @ 7f203b66, card igla-coder-gpu @ 510a22c7; igla cell 9: tern-tc-500m grounded, then reviewed (Closes #10420) #10428 grounds LIVE[8] and REVIEWED[8], review 5563369440 on the final head; merged 82bd5f4. The card now carries HumanEval and a partial t27 bench, HF commit 945bf79, igla-coder-gpu#45)
• cell 26: the model computing from DDR3 on the AX7203, reviewed